libyuv: update to r1305

MIPS build fixes https://code.google.com/p/webm/issues/detail?id=957 Change-Id: I9d53900af36d783c369b5dff27a7479cb94fd16b
author: Johann <johannkoenig@google.com> 2015-03-02 15:19:19 -0800
committer: Johann <johannkoenig@google.com> 2015-03-02 15:19:31 -0800
commit: 223bf29307a1981940dbc224a0b1279a187eca6c (patch)
tree: 438d592b79c8f70ef5ea5c0c83cc1c02484c19d3 /third_party/libyuv/source
parent: 1790d45252efb29903baae3d1776bb24cee76558 (diff)
download: libvpx-223bf29307a1981940dbc224a0b1279a187eca6c.tar
libvpx-223bf29307a1981940dbc224a0b1279a187eca6c.tar.gz
libvpx-223bf29307a1981940dbc224a0b1279a187eca6c.tar.bz2
libvpx-223bf29307a1981940dbc224a0b1279a187eca6c.zip
34 files changed, 7182 insertions, 9972 deletions
diff --git a/third_party/libyuv/source/compare.cc b/third_party/libyuv/source/compare.cc
index dc715e019..f84a08ee6 100644
--- a/third_party/libyuv/source/compare.cc
+++ b/third_party/libyuv/source/compare.cc
@@ -19,6 +19,7 @@
 #include "libyuv/basic_types.h"
 #include "libyuv/cpu_id.h"
 #include "libyuv/row.h"
+#include "libyuv/video_common.h"
 
 #ifdef __cplusplus
 namespace libyuv {
@@ -78,6 +79,54 @@ uint32 HashDjb2(const uint8* src, uint64 count, uint32 seed) {
   return seed;
 }
 
+static uint32 ARGBDetectRow_C(const uint8* argb, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    if (argb[0] != 255) {  // First byte is not Alpha of 255, so not ARGB.
+      return FOURCC_BGRA;
+    }
+    if (argb[3] != 255) {  // 4th byte is not Alpha of 255, so not BGRA.
+      return FOURCC_ARGB;
+    }
+    if (argb[4] != 255) {  // Second pixel first byte is not Alpha of 255.
+      return FOURCC_BGRA;
+    }
+    if (argb[7] != 255) {  // Second pixel 4th byte is not Alpha of 255.
+      return FOURCC_ARGB;
+    }
+    argb += 8;
+  }
+  if (width & 1) {
+    if (argb[0] != 255) {  // First byte is not Alpha of 255, so not ARGB.
+      return FOURCC_BGRA;
+    }
+    if (argb[3] != 255) {  // 4th byte is not Alpha of 255, so not BGRA.
+      return FOURCC_ARGB;
+    }
+  }
+  return 0;
+}
+
+// Scan an opaque argb image and return fourcc based on alpha offset.
+// Returns FOURCC_ARGB, FOURCC_BGRA, or 0 if unknown.
+LIBYUV_API
+uint32 ARGBDetect(const uint8* argb, int stride_argb, int width, int height) {
+  uint32 fourcc = 0;
+  int h;
+
+  // Coalesce rows.
+  if (stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    stride_argb = 0;
+  }
+  for (h = 0; h < height && fourcc == 0; ++h) {
+    fourcc = ARGBDetectRow_C(argb, width);
+    argb += stride_argb;
+  }
+  return fourcc;
+}
+
 uint32 SumSquareError_C(const uint8* src_a, const uint8* src_b, int count);
 #if !defined(LIBYUV_DISABLE_NEON) && \
     (defined(__ARM_NEON__) || defined(LIBYUV_NEON) || defined(__aarch64__))
@@ -114,8 +163,7 @@ uint64 ComputeSumSquareError(const uint8* src_a, const uint8* src_b,
   }
 #endif
 #if defined(HAS_SUMSQUAREERROR_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) &&
-      IS_ALIGNED(src_a, 16) && IS_ALIGNED(src_b, 16)) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     // Note only used for multiples of 16 so count is not checked.
     SumSquareError = SumSquareError_SSE2;
   }
diff --git a/third_party/libyuv/source/compare_neon.cc b/third_party/libyuv/source/compare_neon.cc
index 55052c0ee..ef006ec41 100644
--- a/third_party/libyuv/source/compare_neon.cc
+++ b/third_party/libyuv/source/compare_neon.cc
@@ -16,7 +16,8 @@ namespace libyuv {
 extern "C" {
 #endif
 
-#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__)
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \
+    !defined(__aarch64__)
 
 uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count) {
   volatile uint32 sse;
@@ -56,46 +57,7 @@ uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count) {
   return sse;
 }
 
-#elif !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
-
-uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count) {
-  volatile uint32 sse;
-  asm volatile (
-    "eor        v16.16b, v16.16b, v16.16b      \n"
-    "eor        v18.16b, v18.16b, v18.16b      \n"
-    "eor        v17.16b, v17.16b, v17.16b      \n"
-    "eor        v19.16b, v19.16b, v19.16b      \n"
-
-    ".p2align  2                               \n"
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.16b}, [%0], #16            \n"
-    MEMACCESS(1)
-    "ld1        {v1.16b}, [%1], #16            \n"
-    "subs       %2, %2, #16                    \n"
-    "usubl      v2.8h, v0.8b, v1.8b            \n"
-    "usubl2     v3.8h, v0.16b, v1.16b          \n"
-    "smlal      v16.4s, v2.4h, v2.4h           \n"
-    "smlal      v17.4s, v3.4h, v3.4h           \n"
-    "smlal2     v18.4s, v2.8h, v2.8h           \n"
-    "smlal2     v19.4s, v3.8h, v3.8h           \n"
-    "bgt        1b                             \n"
-
-    "add        v16.4s, v16.4s, v17.4s         \n"
-    "add        v18.4s, v18.4s, v19.4s         \n"
-    "add        v19.4s, v16.4s, v18.4s         \n"
-    "addv       s0, v19.4s                     \n"
-    "fmov       %w3, s0                        \n"
-    : "+r"(src_a),
-      "+r"(src_b),
-      "+r"(count),
-      "=r"(sse)
-    :
-    : "cc", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19");
-  return sse;
-}
-
-#endif  // __ARM_NEON__
+#endif  // defined(__ARM_NEON__) && !defined(__aarch64__)
 
 #ifdef __cplusplus
 }  // extern "C"
diff --git a/third_party/libyuv/source/compare_neon64.cc b/third_party/libyuv/source/compare_neon64.cc
new file mode 100644
index 000000000..cc078f84c
--- /dev/null
+++ b/third_party/libyuv/source/compare_neon64.cc
@@ -0,0 +1,63 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count) {
+  volatile uint32 sse;
+  asm volatile (
+    "eor        v16.16b, v16.16b, v16.16b      \n"
+    "eor        v18.16b, v18.16b, v18.16b      \n"
+    "eor        v17.16b, v17.16b, v17.16b      \n"
+    "eor        v19.16b, v19.16b, v19.16b      \n"
+
+    ".p2align  2                               \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"
+    MEMACCESS(1)
+    "ld1        {v1.16b}, [%1], #16            \n"
+    "subs       %2, %2, #16                    \n"
+    "usubl      v2.8h, v0.8b, v1.8b            \n"
+    "usubl2     v3.8h, v0.16b, v1.16b          \n"
+    "smlal      v16.4s, v2.4h, v2.4h           \n"
+    "smlal      v17.4s, v3.4h, v3.4h           \n"
+    "smlal2     v18.4s, v2.8h, v2.8h           \n"
+    "smlal2     v19.4s, v3.8h, v3.8h           \n"
+    "b.gt       1b                             \n"
+
+    "add        v16.4s, v16.4s, v17.4s         \n"
+    "add        v18.4s, v18.4s, v19.4s         \n"
+    "add        v19.4s, v16.4s, v18.4s         \n"
+    "addv       s0, v19.4s                     \n"
+    "fmov       %w3, s0                        \n"
+    : "+r"(src_a),
+      "+r"(src_b),
+      "+r"(count),
+      "=r"(sse)
+    :
+    : "cc", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19");
+  return sse;
+}
+
+#endif  // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/libyuv/source/compare_posix.cc b/third_party/libyuv/source/compare_posix.cc
index ac361190e..247cb33bb 100644
--- a/third_party/libyuv/source/compare_posix.cc
+++ b/third_party/libyuv/source/compare_posix.cc
@@ -25,11 +25,10 @@ uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
     "pxor      %%xmm5,%%xmm5                   \n"
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm1         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
     "lea       " MEMLEA(0x10, 0) ",%0          \n"
-    "movdqa    " MEMACCESS(1) ",%%xmm2         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm2         \n"
     "lea       " MEMLEA(0x10, 1) ",%1          \n"
-    "sub       $0x10,%2                        \n"
     "movdqa    %%xmm1,%%xmm3                   \n"
     "psubusb   %%xmm2,%%xmm1                   \n"
     "psubusb   %%xmm3,%%xmm2                   \n"
@@ -41,6 +40,7 @@ uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
     "pmaddwd   %%xmm2,%%xmm2                   \n"
     "paddd     %%xmm1,%%xmm0                   \n"
     "paddd     %%xmm2,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
     "jg        1b                              \n"
 
     "pshufd    $0xee,%%xmm0,%%xmm1             \n"
@@ -53,11 +53,7 @@ uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
     "+r"(src_b),      // %1
     "+r"(count),      // %2
     "=g"(sse)         // %3
-  :
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-#endif
+  :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
   );  // NOLINT
   return sse;
 }
@@ -124,13 +120,13 @@ uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
     "pmulld    %%xmm5,%%xmm1                   \n"
     "paddd     %%xmm4,%%xmm3                   \n"
     "paddd     %%xmm2,%%xmm1                   \n"
-    "sub       $0x10,%1                        \n"
     "paddd     %%xmm3,%%xmm1                   \n"
     "pshufd    $0xe,%%xmm1,%%xmm2              \n"
     "paddd     %%xmm2,%%xmm1                   \n"
     "pshufd    $0x1,%%xmm1,%%xmm2              \n"
     "paddd     %%xmm2,%%xmm1                   \n"
     "paddd     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%1                        \n"
     "jg        1b                              \n"
     "movd      %%xmm0,%3                       \n"
   : "+r"(src),        // %0
@@ -143,9 +139,7 @@ uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
     "m"(kHashMul2),   // %7
     "m"(kHashMul3)    // %8
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-#endif
   );  // NOLINT
   return hash;
 }
diff --git a/third_party/libyuv/source/compare_win.cc b/third_party/libyuv/source/compare_win.cc
index 99831651f..e99009a21 100644
--- a/third_party/libyuv/source/compare_win.cc
+++ b/third_party/libyuv/source/compare_win.cc
@@ -27,13 +27,11 @@ uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
     pxor       xmm0, xmm0
     pxor       xmm5, xmm5
 
-    align      4
   wloop:
-    movdqa     xmm1, [eax]
+    movdqu     xmm1, [eax]
     lea        eax,  [eax + 16]
-    movdqa     xmm2, [edx]
+    movdqu     xmm2, [edx]
     lea        edx,  [edx + 16]
-    sub        ecx, 16
     movdqa     xmm3, xmm1  // abs trick
     psubusb    xmm1, xmm2
     psubusb    xmm2, xmm3
@@ -45,6 +43,7 @@ uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
     pmaddwd    xmm2, xmm2
     paddd      xmm0, xmm1
     paddd      xmm0, xmm2
+    sub        ecx, 16
     jg         wloop
 
     pshufd     xmm1, xmm0, 0xee
@@ -70,12 +69,10 @@ uint32 SumSquareError_AVX2(const uint8* src_a, const uint8* src_b, int count) {
     vpxor      ymm5, ymm5, ymm5  // constant 0 for unpck
     sub        edx, eax
 
-    align      4
   wloop:
     vmovdqu    ymm1, [eax]
     vmovdqu    ymm2, [eax + edx]
     lea        eax,  [eax + 32]
-    sub        ecx, 32
     vpsubusb   ymm3, ymm1, ymm2  // abs difference trick
     vpsubusb   ymm2, ymm2, ymm1
     vpor       ymm1, ymm2, ymm3
@@ -85,6 +82,7 @@ uint32 SumSquareError_AVX2(const uint8* src_a, const uint8* src_b, int count) {
     vpmaddwd   ymm1, ymm1, ymm1
     vpaddd     ymm0, ymm0, ymm1
     vpaddd     ymm0, ymm0, ymm2
+    sub        ecx, 32
     jg         wloop
 
     vpshufd    ymm1, ymm0, 0xee  // 3, 2 + 1, 0 both lanes.
@@ -145,7 +143,6 @@ uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
     pxor       xmm7, xmm7        // constant 0 for unpck
     movdqa     xmm6, kHash16x33
 
-    align      4
   wloop:
     movdqu     xmm1, [eax]       // src[0-15]
     lea        eax, [eax + 16]
@@ -170,7 +167,6 @@ uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
     pmulld(0xcd)                 // pmulld     xmm1, xmm5
     paddd      xmm3, xmm4        // add 16 results
     paddd      xmm1, xmm2
-    sub        ecx, 16
     paddd      xmm1, xmm3
 
     pshufd     xmm2, xmm1, 0x0e  // upper 2 dwords
@@ -178,6 +174,7 @@ uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
     pshufd     xmm2, xmm1, 0x01
     paddd      xmm1, xmm2
     paddd      xmm0, xmm1
+    sub        ecx, 16
     jg         wloop
 
     movd       eax, xmm0         // return hash
@@ -195,7 +192,6 @@ uint32 HashDjb2_AVX2(const uint8* src, int count, uint32 seed) {
     movd       xmm0, [esp + 12]  // seed
     movdqa     xmm6, kHash16x33
 
-    align      4
   wloop:
     vpmovzxbd  xmm3, dword ptr [eax]  // src[0-3]
     pmulld     xmm0, xmm6  // hash *= 33 ^ 16
@@ -209,13 +205,13 @@ uint32 HashDjb2_AVX2(const uint8* src, int count, uint32 seed) {
     pmulld     xmm1, kHashMul3
     paddd      xmm3, xmm4        // add 16 results
     paddd      xmm1, xmm2
-    sub        ecx, 16
     paddd      xmm1, xmm3
     pshufd     xmm2, xmm1, 0x0e  // upper 2 dwords
     paddd      xmm1, xmm2
     pshufd     xmm2, xmm1, 0x01
     paddd      xmm1, xmm2
     paddd      xmm0, xmm1
+    sub        ecx, 16
     jg         wloop
 
     movd       eax, xmm0         // return hash
diff --git a/third_party/libyuv/source/convert.cc b/third_party/libyuv/source/convert.cc
index a8e294f47..41696c18f 100644
--- a/third_party/libyuv/source/convert.cc
+++ b/third_party/libyuv/source/convert.cc
@@ -188,17 +188,14 @@ static void CopyPlane2(const uint8* src, int src_stride_0, int src_stride_1,
                        int width, int height) {
   int y;
   void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
-#if defined(HAS_COPYROW_X86)
-  if (TestCpuFlag(kCpuHasX86) && IS_ALIGNED(width, 4)) {
-    CopyRow = CopyRow_X86;
+#if defined(HAS_COPYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    CopyRow = IS_ALIGNED(width, 32) ? CopyRow_SSE2 : CopyRow_Any_SSE2;
   }
 #endif
-#if defined(HAS_COPYROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 32) &&
-      IS_ALIGNED(src, 16) &&
-      IS_ALIGNED(src_stride_0, 16) && IS_ALIGNED(src_stride_1, 16) &&
-      IS_ALIGNED(dst, 16) && IS_ALIGNED(dst_stride, 16)) {
-    CopyRow = CopyRow_SSE2;
+#if defined(HAS_COPYROW_AVX)
+  if (TestCpuFlag(kCpuHasAVX)) {
+    CopyRow = IS_ALIGNED(width, 64) ? CopyRow_AVX : CopyRow_Any_AVX;
   }
 #endif
 #if defined(HAS_COPYROW_ERMS)
@@ -207,8 +204,8 @@ static void CopyPlane2(const uint8* src, int src_stride_0, int src_stride_1,
   }
 #endif
 #if defined(HAS_COPYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 32)) {
-    CopyRow = CopyRow_NEON;
+  if (TestCpuFlag(kCpuHasNEON)) {
+    CopyRow = IS_ALIGNED(width, 32) ? CopyRow_NEON : CopyRow_Any_NEON;
   }
 #endif
 #if defined(HAS_COPYROW_MIPS)
@@ -283,20 +280,15 @@ static int X420ToI420(const uint8* src_y,
     src_stride_uv = dst_stride_u = dst_stride_v = 0;
   }
 #if defined(HAS_SPLITUVROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && halfwidth >= 16) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     SplitUVRow = SplitUVRow_Any_SSE2;
     if (IS_ALIGNED(halfwidth, 16)) {
-      SplitUVRow = SplitUVRow_Unaligned_SSE2;
-      if (IS_ALIGNED(src_uv, 16) && IS_ALIGNED(src_stride_uv, 16) &&
-          IS_ALIGNED(dst_u, 16) && IS_ALIGNED(dst_stride_u, 16) &&
-          IS_ALIGNED(dst_v, 16) && IS_ALIGNED(dst_stride_v, 16)) {
-        SplitUVRow = SplitUVRow_SSE2;
-      }
+      SplitUVRow = SplitUVRow_SSE2;
     }
   }
 #endif
 #if defined(HAS_SPLITUVROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && halfwidth >= 32) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     SplitUVRow = SplitUVRow_Any_AVX2;
     if (IS_ALIGNED(halfwidth, 32)) {
       SplitUVRow = SplitUVRow_AVX2;
@@ -304,7 +296,7 @@ static int X420ToI420(const uint8* src_y,
   }
 #endif
 #if defined(HAS_SPLITUVROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && halfwidth >= 16) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     SplitUVRow = SplitUVRow_Any_NEON;
     if (IS_ALIGNED(halfwidth, 16)) {
       SplitUVRow = SplitUVRow_NEON;
@@ -312,15 +304,13 @@ static int X420ToI420(const uint8* src_y,
   }
 #endif
 #if defined(HAS_SPLITUVROW_MIPS_DSPR2)
-  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && halfwidth >= 16) {
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(src_uv, 4) && IS_ALIGNED(src_stride_uv, 4) &&
+      IS_ALIGNED(dst_u, 4) && IS_ALIGNED(dst_stride_u, 4) &&
+      IS_ALIGNED(dst_v, 4) && IS_ALIGNED(dst_stride_v, 4)) {
     SplitUVRow = SplitUVRow_Any_MIPS_DSPR2;
     if (IS_ALIGNED(halfwidth, 16)) {
-      SplitUVRow = SplitUVRow_Unaligned_MIPS_DSPR2;
-      if (IS_ALIGNED(src_uv, 4) && IS_ALIGNED(src_stride_uv, 4) &&
-          IS_ALIGNED(dst_u, 4) && IS_ALIGNED(dst_stride_u, 4) &&
-          IS_ALIGNED(dst_v, 4) && IS_ALIGNED(dst_stride_v, 4)) {
-        SplitUVRow = SplitUVRow_MIPS_DSPR2;
-      }
+      SplitUVRow = SplitUVRow_MIPS_DSPR2;
     }
   }
 #endif
@@ -391,125 +381,6 @@ int M420ToI420(const uint8* src_m420, int src_stride_m420,
                     width, height);
 }
 
-// Convert Q420 to I420.
-// Format is rows of YY/YUYV
-LIBYUV_API
-int Q420ToI420(const uint8* src_y, int src_stride_y,
-               const uint8* src_yuy2, int src_stride_yuy2,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
-  int y;
-  int halfheight;
-  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
-  void (*YUY2ToUV422Row)(const uint8* src_yuy2, uint8* dst_u, uint8* dst_v,
-      int pix) = YUY2ToUV422Row_C;
-  void (*YUY2ToYRow)(const uint8* src_yuy2, uint8* dst_y, int pix) =
-      YUY2ToYRow_C;
-  if (!src_y || !src_yuy2 ||
-      !dst_y || !dst_u || !dst_v ||
-      width <= 0 || height == 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (height < 0) {
-    height = -height;
-    halfheight = (height + 1) >> 1;
-    dst_y = dst_y + (height - 1) * dst_stride_y;
-    dst_u = dst_u + (halfheight - 1) * dst_stride_u;
-    dst_v = dst_v + (halfheight - 1) * dst_stride_v;
-    dst_stride_y = -dst_stride_y;
-    dst_stride_u = -dst_stride_u;
-    dst_stride_v = -dst_stride_v;
-  }
-  // CopyRow for rows of just Y in Q420 copied to Y plane of I420.
-#if defined(HAS_COPYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 32)) {
-    CopyRow = CopyRow_NEON;
-  }
-#endif
-#if defined(HAS_COPYROW_X86)
-  if (IS_ALIGNED(width, 4)) {
-    CopyRow = CopyRow_X86;
-  }
-#endif
-#if defined(HAS_COPYROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 32) &&
-      IS_ALIGNED(src_y, 16) && IS_ALIGNED(src_stride_y, 16) &&
-      IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-    CopyRow = CopyRow_SSE2;
-  }
-#endif
-#if defined(HAS_COPYROW_ERMS)
-  if (TestCpuFlag(kCpuHasERMS)) {
-    CopyRow = CopyRow_ERMS;
-  }
-#endif
-#if defined(HAS_COPYROW_MIPS)
-  if (TestCpuFlag(kCpuHasMIPS)) {
-    CopyRow = CopyRow_MIPS;
-  }
-#endif
-
-#if defined(HAS_YUY2TOYROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
-    YUY2ToUV422Row = YUY2ToUV422Row_Any_SSE2;
-    YUY2ToYRow = YUY2ToYRow_Any_SSE2;
-    if (IS_ALIGNED(width, 16)) {
-      YUY2ToUV422Row = YUY2ToUV422Row_Unaligned_SSE2;
-      YUY2ToYRow = YUY2ToYRow_Unaligned_SSE2;
-      if (IS_ALIGNED(src_yuy2, 16) && IS_ALIGNED(src_stride_yuy2, 16)) {
-        YUY2ToUV422Row = YUY2ToUV422Row_SSE2;
-        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-          YUY2ToYRow = YUY2ToYRow_SSE2;
-        }
-      }
-    }
-  }
-#endif
-#if defined(HAS_YUY2TOYROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
-    YUY2ToUV422Row = YUY2ToUV422Row_Any_AVX2;
-    YUY2ToYRow = YUY2ToYRow_Any_AVX2;
-    if (IS_ALIGNED(width, 32)) {
-      YUY2ToUV422Row = YUY2ToUV422Row_AVX2;
-      YUY2ToYRow = YUY2ToYRow_AVX2;
-    }
-  }
-#endif
-#if defined(HAS_YUY2TOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
-    YUY2ToYRow = YUY2ToYRow_Any_NEON;
-    if (width >= 16) {
-      YUY2ToUV422Row = YUY2ToUV422Row_Any_NEON;
-    }
-    if (IS_ALIGNED(width, 16)) {
-      YUY2ToYRow = YUY2ToYRow_NEON;
-      YUY2ToUV422Row = YUY2ToUV422Row_NEON;
-    }
-  }
-#endif
-
-  for (y = 0; y < height - 1; y += 2) {
-    CopyRow(src_y, dst_y, width);
-    src_y += src_stride_y;
-    dst_y += dst_stride_y;
-
-    YUY2ToUV422Row(src_yuy2, dst_u, dst_v, width);
-    YUY2ToYRow(src_yuy2, dst_y, width);
-    src_yuy2 += src_stride_yuy2;
-    dst_y += dst_stride_y;
-    dst_u += dst_stride_u;
-    dst_v += dst_stride_v;
-  }
-  if (height & 1) {
-    CopyRow(src_y, dst_y, width);
-    YUY2ToUV422Row(src_yuy2, dst_u, dst_v, width);
-  }
-  return 0;
-}
-
 // Convert YUY2 to I420.
 LIBYUV_API
 int YUY2ToI420(const uint8* src_yuy2, int src_stride_yuy2,
@@ -529,23 +400,17 @@ int YUY2ToI420(const uint8* src_yuy2, int src_stride_yuy2,
     src_stride_yuy2 = -src_stride_yuy2;
   }
 #if defined(HAS_YUY2TOYROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     YUY2ToUVRow = YUY2ToUVRow_Any_SSE2;
     YUY2ToYRow = YUY2ToYRow_Any_SSE2;
     if (IS_ALIGNED(width, 16)) {
-      YUY2ToUVRow = YUY2ToUVRow_Unaligned_SSE2;
-      YUY2ToYRow = YUY2ToYRow_Unaligned_SSE2;
-      if (IS_ALIGNED(src_yuy2, 16) && IS_ALIGNED(src_stride_yuy2, 16)) {
-        YUY2ToUVRow = YUY2ToUVRow_SSE2;
-        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-          YUY2ToYRow = YUY2ToYRow_SSE2;
-        }
-      }
+      YUY2ToUVRow = YUY2ToUVRow_SSE2;
+      YUY2ToYRow = YUY2ToYRow_SSE2;
     }
   }
 #endif
 #if defined(HAS_YUY2TOYROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     YUY2ToUVRow = YUY2ToUVRow_Any_AVX2;
     YUY2ToYRow = YUY2ToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
@@ -555,11 +420,9 @@ int YUY2ToI420(const uint8* src_yuy2, int src_stride_yuy2,
   }
 #endif
 #if defined(HAS_YUY2TOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     YUY2ToYRow = YUY2ToYRow_Any_NEON;
-    if (width >= 16) {
-      YUY2ToUVRow = YUY2ToUVRow_Any_NEON;
-    }
+    YUY2ToUVRow = YUY2ToUVRow_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
       YUY2ToYRow = YUY2ToYRow_NEON;
       YUY2ToUVRow = YUY2ToUVRow_NEON;
@@ -602,23 +465,17 @@ int UYVYToI420(const uint8* src_uyvy, int src_stride_uyvy,
     src_stride_uyvy = -src_stride_uyvy;
   }
 #if defined(HAS_UYVYTOYROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     UYVYToUVRow = UYVYToUVRow_Any_SSE2;
     UYVYToYRow = UYVYToYRow_Any_SSE2;
     if (IS_ALIGNED(width, 16)) {
-      UYVYToUVRow = UYVYToUVRow_Unaligned_SSE2;
-      UYVYToYRow = UYVYToYRow_Unaligned_SSE2;
-      if (IS_ALIGNED(src_uyvy, 16) && IS_ALIGNED(src_stride_uyvy, 16)) {
-        UYVYToUVRow = UYVYToUVRow_SSE2;
-        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-          UYVYToYRow = UYVYToYRow_SSE2;
-        }
-      }
+      UYVYToUVRow = UYVYToUVRow_SSE2;
+      UYVYToYRow = UYVYToYRow_SSE2;
     }
   }
 #endif
 #if defined(HAS_UYVYTOYROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     UYVYToUVRow = UYVYToUVRow_Any_AVX2;
     UYVYToYRow = UYVYToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
@@ -628,11 +485,9 @@ int UYVYToI420(const uint8* src_uyvy, int src_stride_uyvy,
   }
 #endif
 #if defined(HAS_UYVYTOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     UYVYToYRow = UYVYToYRow_Any_NEON;
-    if (width >= 16) {
-      UYVYToUVRow = UYVYToUVRow_Any_NEON;
-    }
+    UYVYToUVRow = UYVYToUVRow_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
       UYVYToYRow = UYVYToYRow_NEON;
       UYVYToUVRow = UYVYToUVRow_NEON;
@@ -680,23 +535,17 @@ int ARGBToI420(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = -src_stride_argb;
   }
 #if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_Unaligned_SSSE3;
-      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
-        ARGBToUVRow = ARGBToUVRow_SSSE3;
-        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-          ARGBToYRow = ARGBToYRow_SSSE3;
-        }
-      }
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+      ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
 #if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     ARGBToUVRow = ARGBToUVRow_Any_AVX2;
     ARGBToYRow = ARGBToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
@@ -706,7 +555,7 @@ int ARGBToI420(const uint8* src_argb, int src_stride_argb,
   }
 #endif
 #if defined(HAS_ARGBTOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGBToYRow = ARGBToYRow_NEON;
@@ -714,7 +563,7 @@ int ARGBToI420(const uint8* src_argb, int src_stride_argb,
   }
 #endif
 #if defined(HAS_ARGBTOUVROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToUVRow = ARGBToUVRow_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
       ARGBToUVRow = ARGBToUVRow_NEON;
@@ -761,34 +610,31 @@ int BGRAToI420(const uint8* src_bgra, int src_stride_bgra,
     src_bgra = src_bgra + (height - 1) * src_stride_bgra;
     src_stride_bgra = -src_stride_bgra;
   }
-#if defined(HAS_BGRATOYROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+#if defined(HAS_BGRATOYROW_SSSE3) && defined(HAS_BGRATOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     BGRAToUVRow = BGRAToUVRow_Any_SSSE3;
     BGRAToYRow = BGRAToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      BGRAToUVRow = BGRAToUVRow_Unaligned_SSSE3;
-      BGRAToYRow = BGRAToYRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(src_bgra, 16) && IS_ALIGNED(src_stride_bgra, 16)) {
-        BGRAToUVRow = BGRAToUVRow_SSSE3;
-        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-          BGRAToYRow = BGRAToYRow_SSSE3;
-        }
-      }
+      BGRAToUVRow = BGRAToUVRow_SSSE3;
+      BGRAToYRow = BGRAToYRow_SSSE3;
     }
   }
-#elif defined(HAS_BGRATOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_BGRATOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     BGRAToYRow = BGRAToYRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       BGRAToYRow = BGRAToYRow_NEON;
     }
-    if (width >= 16) {
+  }
+#endif
+#if defined(HAS_BGRATOUVROW_NEON)
+    if (TestCpuFlag(kCpuHasNEON)) {
       BGRAToUVRow = BGRAToUVRow_Any_NEON;
       if (IS_ALIGNED(width, 16)) {
         BGRAToUVRow = BGRAToUVRow_NEON;
       }
     }
-  }
 #endif
 
   for (y = 0; y < height - 1; y += 2) {
@@ -830,32 +676,29 @@ int ABGRToI420(const uint8* src_abgr, int src_stride_abgr,
     src_abgr = src_abgr + (height - 1) * src_stride_abgr;
     src_stride_abgr = -src_stride_abgr;
   }
-#if defined(HAS_ABGRTOYROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+#if defined(HAS_ABGRTOYROW_SSSE3) && defined(HAS_ABGRTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ABGRToUVRow = ABGRToUVRow_Any_SSSE3;
     ABGRToYRow = ABGRToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ABGRToUVRow = ABGRToUVRow_Unaligned_SSSE3;
-      ABGRToYRow = ABGRToYRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(src_abgr, 16) && IS_ALIGNED(src_stride_abgr, 16)) {
-        ABGRToUVRow = ABGRToUVRow_SSSE3;
-        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-          ABGRToYRow = ABGRToYRow_SSSE3;
-        }
-      }
+      ABGRToUVRow = ABGRToUVRow_SSSE3;
+      ABGRToYRow = ABGRToYRow_SSSE3;
     }
   }
-#elif defined(HAS_ABGRTOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_ABGRTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     ABGRToYRow = ABGRToYRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ABGRToYRow = ABGRToYRow_NEON;
     }
-    if (width >= 16) {
-      ABGRToUVRow = ABGRToUVRow_Any_NEON;
-      if (IS_ALIGNED(width, 16)) {
-        ABGRToUVRow = ABGRToUVRow_NEON;
-      }
+  }
+#endif
+#if defined(HAS_ABGRTOUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ABGRToUVRow = ABGRToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToUVRow = ABGRToUVRow_NEON;
     }
   }
 #endif
@@ -899,32 +742,29 @@ int RGBAToI420(const uint8* src_rgba, int src_stride_rgba,
     src_rgba = src_rgba + (height - 1) * src_stride_rgba;
     src_stride_rgba = -src_stride_rgba;
   }
-#if defined(HAS_RGBATOYROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+#if defined(HAS_RGBATOYROW_SSSE3) && defined(HAS_RGBATOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     RGBAToUVRow = RGBAToUVRow_Any_SSSE3;
     RGBAToYRow = RGBAToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      RGBAToUVRow = RGBAToUVRow_Unaligned_SSSE3;
-      RGBAToYRow = RGBAToYRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(src_rgba, 16) && IS_ALIGNED(src_stride_rgba, 16)) {
-        RGBAToUVRow = RGBAToUVRow_SSSE3;
-        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-          RGBAToYRow = RGBAToYRow_SSSE3;
-        }
-      }
+      RGBAToUVRow = RGBAToUVRow_SSSE3;
+      RGBAToYRow = RGBAToYRow_SSSE3;
     }
   }
-#elif defined(HAS_RGBATOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_RGBATOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     RGBAToYRow = RGBAToYRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       RGBAToYRow = RGBAToYRow_NEON;
     }
-    if (width >= 16) {
-      RGBAToUVRow = RGBAToUVRow_Any_NEON;
-      if (IS_ALIGNED(width, 16)) {
-        RGBAToUVRow = RGBAToUVRow_NEON;
-      }
+  }
+#endif
+#if defined(HAS_RGBATOUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RGBAToUVRow = RGBAToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      RGBAToUVRow = RGBAToUVRow_NEON;
     }
   }
 #endif
@@ -978,22 +818,23 @@ int RGB24ToI420(const uint8* src_rgb24, int src_stride_rgb24,
   }
 
 #if defined(HAS_RGB24TOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     RGB24ToYRow = RGB24ToYRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       RGB24ToYRow = RGB24ToYRow_NEON;
     }
-    if (width >= 16) {
-      RGB24ToUVRow = RGB24ToUVRow_Any_NEON;
-      if (IS_ALIGNED(width, 16)) {
-        RGB24ToUVRow = RGB24ToUVRow_NEON;
-      }
+  }
+#endif
+#if defined(HAS_RGB24TOUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RGB24ToUVRow = RGB24ToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToUVRow = RGB24ToUVRow_NEON;
     }
   }
-#else  // HAS_RGB24TOYROW_NEON
-
+#endif
 #if defined(HAS_RGB24TOARGBROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     RGB24ToARGBRow = RGB24ToARGBRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
       RGB24ToARGBRow = RGB24ToARGBRow_SSSE3;
@@ -1001,7 +842,7 @@ int RGB24ToI420(const uint8* src_rgb24, int src_stride_rgb24,
   }
 #endif
 #if defined(HAS_ARGBTOUVROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
       ARGBToUVRow = ARGBToUVRow_SSSE3;
@@ -1009,17 +850,13 @@ int RGB24ToI420(const uint8* src_rgb24, int src_stride_rgb24,
   }
 #endif
 #if defined(HAS_ARGBTOUVROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-        ARGBToYRow = ARGBToYRow_SSSE3;
-      }
+      ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif  // HAS_ARGBTOUVROW_SSSE3
-#endif  // HAS_RGB24TOYROW_NEON
 
   {
 #if !defined(HAS_RGB24TOYROW_NEON)
@@ -1095,22 +932,23 @@ int RAWToI420(const uint8* src_raw, int src_stride_raw,
   }
 
 #if defined(HAS_RAWTOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     RAWToYRow = RAWToYRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       RAWToYRow = RAWToYRow_NEON;
     }
-    if (width >= 16) {
-      RAWToUVRow = RAWToUVRow_Any_NEON;
-      if (IS_ALIGNED(width, 16)) {
-        RAWToUVRow = RAWToUVRow_NEON;
-      }
+  }
+#endif
+#if defined(HAS_RAWTOUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RAWToUVRow = RAWToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToUVRow = RAWToUVRow_NEON;
     }
   }
-#else  // HAS_RAWTOYROW_NEON
-
+#endif
 #if defined(HAS_RAWTOARGBROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     RAWToARGBRow = RAWToARGBRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
       RAWToARGBRow = RAWToARGBRow_SSSE3;
@@ -1118,7 +956,7 @@ int RAWToI420(const uint8* src_raw, int src_stride_raw,
   }
 #endif
 #if defined(HAS_ARGBTOUVROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
       ARGBToUVRow = ARGBToUVRow_SSSE3;
@@ -1126,17 +964,13 @@ int RAWToI420(const uint8* src_raw, int src_stride_raw,
   }
 #endif
 #if defined(HAS_ARGBTOUVROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-        ARGBToYRow = ARGBToYRow_SSSE3;
-      }
+      ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif  // HAS_ARGBTOUVROW_SSSE3
-#endif  // HAS_RAWTOYROW_NEON
 
   {
     // Allocate 2 rows of ARGB.
@@ -1210,22 +1044,20 @@ int RGB565ToI420(const uint8* src_rgb565, int src_stride_rgb565,
   }
 
 #if defined(HAS_RGB565TOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     RGB565ToYRow = RGB565ToYRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       RGB565ToYRow = RGB565ToYRow_NEON;
     }
-    if (width >= 16) {
-      RGB565ToUVRow = RGB565ToUVRow_Any_NEON;
-      if (IS_ALIGNED(width, 16)) {
-        RGB565ToUVRow = RGB565ToUVRow_NEON;
-      }
+    RGB565ToUVRow = RGB565ToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      RGB565ToUVRow = RGB565ToUVRow_NEON;
     }
   }
 #else  // HAS_RGB565TOYROW_NEON
 
 #if defined(HAS_RGB565TOARGBROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 8) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     RGB565ToARGBRow = RGB565ToARGBRow_Any_SSE2;
     if (IS_ALIGNED(width, 8)) {
       RGB565ToARGBRow = RGB565ToARGBRow_SSE2;
@@ -1233,7 +1065,7 @@ int RGB565ToI420(const uint8* src_rgb565, int src_stride_rgb565,
   }
 #endif
 #if defined(HAS_ARGBTOUVROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
       ARGBToUVRow = ARGBToUVRow_SSSE3;
@@ -1241,13 +1073,10 @@ int RGB565ToI420(const uint8* src_rgb565, int src_stride_rgb565,
   }
 #endif
 #if defined(HAS_ARGBTOUVROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-        ARGBToYRow = ARGBToYRow_SSSE3;
-      }
+      ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif  // HAS_ARGBTOUVROW_SSSE3
@@ -1327,22 +1156,20 @@ int ARGB1555ToI420(const uint8* src_argb1555, int src_stride_argb1555,
   }
 
 #if defined(HAS_ARGB1555TOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGB1555ToYRow = ARGB1555ToYRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGB1555ToYRow = ARGB1555ToYRow_NEON;
     }
-    if (width >= 16) {
-      ARGB1555ToUVRow = ARGB1555ToUVRow_Any_NEON;
-      if (IS_ALIGNED(width, 16)) {
-        ARGB1555ToUVRow = ARGB1555ToUVRow_NEON;
-      }
+    ARGB1555ToUVRow = ARGB1555ToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGB1555ToUVRow = ARGB1555ToUVRow_NEON;
     }
   }
 #else  // HAS_ARGB1555TOYROW_NEON
 
 #if defined(HAS_ARGB1555TOARGBROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 8) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_SSE2;
     if (IS_ALIGNED(width, 8)) {
       ARGB1555ToARGBRow = ARGB1555ToARGBRow_SSE2;
@@ -1350,7 +1177,7 @@ int ARGB1555ToI420(const uint8* src_argb1555, int src_stride_argb1555,
   }
 #endif
 #if defined(HAS_ARGBTOUVROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
       ARGBToUVRow = ARGBToUVRow_SSSE3;
@@ -1358,13 +1185,10 @@ int ARGB1555ToI420(const uint8* src_argb1555, int src_stride_argb1555,
   }
 #endif
 #if defined(HAS_ARGBTOUVROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-        ARGBToYRow = ARGBToYRow_SSSE3;
-      }
+      ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif  // HAS_ARGBTOUVROW_SSSE3
@@ -1445,22 +1269,20 @@ int ARGB4444ToI420(const uint8* src_argb4444, int src_stride_argb4444,
   }
 
 #if defined(HAS_ARGB4444TOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGB4444ToYRow = ARGB4444ToYRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGB4444ToYRow = ARGB4444ToYRow_NEON;
     }
-    if (width >= 16) {
-      ARGB4444ToUVRow = ARGB4444ToUVRow_Any_NEON;
-      if (IS_ALIGNED(width, 16)) {
-        ARGB4444ToUVRow = ARGB4444ToUVRow_NEON;
-      }
+    ARGB4444ToUVRow = ARGB4444ToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGB4444ToUVRow = ARGB4444ToUVRow_NEON;
     }
   }
 #else  // HAS_ARGB4444TOYROW_NEON
 
 #if defined(HAS_ARGB4444TOARGBROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 8) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_SSE2;
     if (IS_ALIGNED(width, 8)) {
       ARGB4444ToARGBRow = ARGB4444ToARGBRow_SSE2;
@@ -1468,7 +1290,7 @@ int ARGB4444ToI420(const uint8* src_argb4444, int src_stride_argb4444,
   }
 #endif
 #if defined(HAS_ARGBTOUVROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
       ARGBToUVRow = ARGBToUVRow_SSSE3;
@@ -1476,13 +1298,10 @@ int ARGB4444ToI420(const uint8* src_argb4444, int src_stride_argb4444,
   }
 #endif
 #if defined(HAS_ARGBTOUVROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-        ARGBToYRow = ARGBToYRow_SSSE3;
-      }
+      ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif  // HAS_ARGBTOUVROW_SSSE3
diff --git a/third_party/libyuv/source/convert_argb.cc b/third_party/libyuv/source/convert_argb.cc
index ac0bc3d15..66f766079 100644
--- a/third_party/libyuv/source/convert_argb.cc
+++ b/third_party/libyuv/source/convert_argb.cc
@@ -11,7 +11,6 @@
 #include "libyuv/convert_argb.h"
 
 #include "libyuv/cpu_id.h"
-#include "libyuv/format_conversion.h"
 #ifdef HAVE_JPEG
 #include "libyuv/mjpeg_decoder.h"
 #endif
@@ -79,17 +78,15 @@ int I444ToARGB(const uint8* src_y, int src_stride_y,
     src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0;
   }
 #if defined(HAS_I444TOARGBROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     I444ToARGBRow = I444ToARGBRow_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
-      I444ToARGBRow = I444ToARGBRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
-        I444ToARGBRow = I444ToARGBRow_SSSE3;
-      }
+      I444ToARGBRow = I444ToARGBRow_SSSE3;
     }
   }
-#elif defined(HAS_I444TOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_I444TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     I444ToARGBRow = I444ToARGBRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       I444ToARGBRow = I444ToARGBRow_NEON;
@@ -141,18 +138,15 @@ int I422ToARGB(const uint8* src_y, int src_stride_y,
     src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0;
   }
 #if defined(HAS_I422TOARGBROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
-      I422ToARGBRow = I422ToARGBRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
-        I422ToARGBRow = I422ToARGBRow_SSSE3;
-      }
+      I422ToARGBRow = I422ToARGBRow_SSSE3;
     }
   }
 #endif
 #if defined(HAS_I422TOARGBROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && width >= 16) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     I422ToARGBRow = I422ToARGBRow_Any_AVX2;
     if (IS_ALIGNED(width, 16)) {
       I422ToARGBRow = I422ToARGBRow_AVX2;
@@ -160,7 +154,7 @@ int I422ToARGB(const uint8* src_y, int src_stride_y,
   }
 #endif
 #if defined(HAS_I422TOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     I422ToARGBRow = I422ToARGBRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       I422ToARGBRow = I422ToARGBRow_NEON;
@@ -221,17 +215,15 @@ int I411ToARGB(const uint8* src_y, int src_stride_y,
     src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0;
   }
 #if defined(HAS_I411TOARGBROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     I411ToARGBRow = I411ToARGBRow_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
-      I411ToARGBRow = I411ToARGBRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
-        I411ToARGBRow = I411ToARGBRow_SSSE3;
-      }
+      I411ToARGBRow = I411ToARGBRow_SSSE3;
     }
   }
-#elif defined(HAS_I411TOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_I411TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     I411ToARGBRow = I411ToARGBRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       I411ToARGBRow = I411ToARGBRow_NEON;
@@ -276,15 +268,23 @@ int I400ToARGB_Reference(const uint8* src_y, int src_stride_y,
     src_stride_y = dst_stride_argb = 0;
   }
 #if defined(HAS_YTOARGBROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 8 &&
-      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     YToARGBRow = YToARGBRow_Any_SSE2;
     if (IS_ALIGNED(width, 8)) {
       YToARGBRow = YToARGBRow_SSE2;
     }
   }
-#elif defined(HAS_YTOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_YTOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    YToARGBRow = YToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      YToARGBRow = YToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_YTOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     YToARGBRow = YToARGBRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       YToARGBRow = YToARGBRow_NEON;
@@ -326,17 +326,15 @@ int I400ToARGB(const uint8* src_y, int src_stride_y,
     src_stride_y = dst_stride_argb = 0;
   }
 #if defined(HAS_I400TOARGBROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 8) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     I400ToARGBRow = I400ToARGBRow_Any_SSE2;
     if (IS_ALIGNED(width, 8)) {
-      I400ToARGBRow = I400ToARGBRow_Unaligned_SSE2;
-      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
-        I400ToARGBRow = I400ToARGBRow_SSE2;
-      }
+      I400ToARGBRow = I400ToARGBRow_SSE2;
     }
   }
-#elif defined(HAS_I400TOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_I400TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     I400ToARGBRow = I400ToARGBRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       I400ToARGBRow = I400ToARGBRow_NEON;
@@ -447,15 +445,15 @@ int RGB24ToARGB(const uint8* src_rgb24, int src_stride_rgb24,
     src_stride_rgb24 = dst_stride_argb = 0;
   }
 #if defined(HAS_RGB24TOARGBROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16 &&
-      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     RGB24ToARGBRow = RGB24ToARGBRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
       RGB24ToARGBRow = RGB24ToARGBRow_SSSE3;
     }
   }
-#elif defined(HAS_RGB24TOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_RGB24TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     RGB24ToARGBRow = RGB24ToARGBRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       RGB24ToARGBRow = RGB24ToARGBRow_NEON;
@@ -497,15 +495,15 @@ int RAWToARGB(const uint8* src_raw, int src_stride_raw,
     src_stride_raw = dst_stride_argb = 0;
   }
 #if defined(HAS_RAWTOARGBROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16 &&
-      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     RAWToARGBRow = RAWToARGBRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
       RAWToARGBRow = RAWToARGBRow_SSSE3;
     }
   }
-#elif defined(HAS_RAWTOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_RAWTOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     RAWToARGBRow = RAWToARGBRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       RAWToARGBRow = RAWToARGBRow_NEON;
@@ -547,15 +545,15 @@ int RGB565ToARGB(const uint8* src_rgb565, int src_stride_rgb565,
     src_stride_rgb565 = dst_stride_argb = 0;
   }
 #if defined(HAS_RGB565TOARGBROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 8 &&
-      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     RGB565ToARGBRow = RGB565ToARGBRow_Any_SSE2;
     if (IS_ALIGNED(width, 8)) {
       RGB565ToARGBRow = RGB565ToARGBRow_SSE2;
     }
   }
-#elif defined(HAS_RGB565TOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_RGB565TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     RGB565ToARGBRow = RGB565ToARGBRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       RGB565ToARGBRow = RGB565ToARGBRow_NEON;
@@ -597,15 +595,15 @@ int ARGB1555ToARGB(const uint8* src_argb1555, int src_stride_argb1555,
     src_stride_argb1555 = dst_stride_argb = 0;
   }
 #if defined(HAS_ARGB1555TOARGBROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 8 &&
-      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_SSE2;
     if (IS_ALIGNED(width, 8)) {
       ARGB1555ToARGBRow = ARGB1555ToARGBRow_SSE2;
     }
   }
-#elif defined(HAS_ARGB1555TOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_ARGB1555TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGB1555ToARGBRow = ARGB1555ToARGBRow_NEON;
@@ -647,15 +645,15 @@ int ARGB4444ToARGB(const uint8* src_argb4444, int src_stride_argb4444,
     src_stride_argb4444 = dst_stride_argb = 0;
   }
 #if defined(HAS_ARGB4444TOARGBROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 8 &&
-      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_SSE2;
     if (IS_ALIGNED(width, 8)) {
       ARGB4444ToARGBRow = ARGB4444ToARGBRow_SSE2;
     }
   }
-#elif defined(HAS_ARGB4444TOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_ARGB4444TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGB4444ToARGBRow = ARGB4444ToARGBRow_NEON;
@@ -693,17 +691,23 @@ int NV12ToARGB(const uint8* src_y, int src_stride_y,
     dst_stride_argb = -dst_stride_argb;
   }
 #if defined(HAS_NV12TOARGBROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     NV12ToARGBRow = NV12ToARGBRow_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
-      NV12ToARGBRow = NV12ToARGBRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
-        NV12ToARGBRow = NV12ToARGBRow_SSSE3;
-      }
+      NV12ToARGBRow = NV12ToARGBRow_SSSE3;
     }
   }
-#elif defined(HAS_NV12TOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_NV12TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      NV12ToARGBRow = NV12ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_NV12TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     NV12ToARGBRow = NV12ToARGBRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       NV12ToARGBRow = NV12ToARGBRow_NEON;
@@ -744,18 +748,23 @@ int NV21ToARGB(const uint8* src_y, int src_stride_y,
     dst_stride_argb = -dst_stride_argb;
   }
 #if defined(HAS_NV21TOARGBROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     NV21ToARGBRow = NV21ToARGBRow_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
-      NV21ToARGBRow = NV21ToARGBRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
-        NV21ToARGBRow = NV21ToARGBRow_SSSE3;
-      }
+      NV21ToARGBRow = NV21ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_NV21TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    NV21ToARGBRow = NV21ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      NV21ToARGBRow = NV21ToARGBRow_AVX2;
     }
   }
 #endif
 #if defined(HAS_NV21TOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     NV21ToARGBRow = NV21ToARGBRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       NV21ToARGBRow = NV21ToARGBRow_NEON;
@@ -795,17 +804,23 @@ int M420ToARGB(const uint8* src_m420, int src_stride_m420,
     dst_stride_argb = -dst_stride_argb;
   }
 #if defined(HAS_NV12TOARGBROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     NV12ToARGBRow = NV12ToARGBRow_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
-      NV12ToARGBRow = NV12ToARGBRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
-        NV12ToARGBRow = NV12ToARGBRow_SSSE3;
-      }
+      NV12ToARGBRow = NV12ToARGBRow_SSSE3;
     }
   }
-#elif defined(HAS_NV12TOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_NV12TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      NV12ToARGBRow = NV12ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_NV12TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     NV12ToARGBRow = NV12ToARGBRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       NV12ToARGBRow = NV12ToARGBRow_NEON;
@@ -852,19 +867,23 @@ int YUY2ToARGB(const uint8* src_yuy2, int src_stride_yuy2,
     src_stride_yuy2 = dst_stride_argb = 0;
   }
 #if defined(HAS_YUY2TOARGBROW_SSSE3)
-  // Posix is 16, Windows is 8.
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     YUY2ToARGBRow = YUY2ToARGBRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      YUY2ToARGBRow = YUY2ToARGBRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(src_yuy2, 16) && IS_ALIGNED(src_stride_yuy2, 16) &&
-          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
-        YUY2ToARGBRow = YUY2ToARGBRow_SSSE3;
-      }
+      YUY2ToARGBRow = YUY2ToARGBRow_SSSE3;
     }
   }
-#elif defined(HAS_YUY2TOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_YUY2TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    YUY2ToARGBRow = YUY2ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToARGBRow = YUY2ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     YUY2ToARGBRow = YUY2ToARGBRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       YUY2ToARGBRow = YUY2ToARGBRow_NEON;
@@ -905,19 +924,23 @@ int UYVYToARGB(const uint8* src_uyvy, int src_stride_uyvy,
     src_stride_uyvy = dst_stride_argb = 0;
   }
 #if defined(HAS_UYVYTOARGBROW_SSSE3)
-  // Posix is 16, Windows is 8.
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     UYVYToARGBRow = UYVYToARGBRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      UYVYToARGBRow = UYVYToARGBRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(src_uyvy, 16) && IS_ALIGNED(src_stride_uyvy, 16) &&
-          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
-        UYVYToARGBRow = UYVYToARGBRow_SSSE3;
-      }
+      UYVYToARGBRow = UYVYToARGBRow_SSSE3;
     }
   }
-#elif defined(HAS_UYVYTOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_UYVYTOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    UYVYToARGBRow = UYVYToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      UYVYToARGBRow = UYVYToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     UYVYToARGBRow = UYVYToARGBRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       UYVYToARGBRow = UYVYToARGBRow_NEON;
@@ -932,6 +955,152 @@ int UYVYToARGB(const uint8* src_uyvy, int src_stride_uyvy,
   return 0;
 }
 
+// Convert J420 to ARGB.
+LIBYUV_API
+int J420ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*J422ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = J422ToARGBRow_C;
+  if (!src_y || !src_u || !src_v || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_J422TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    J422ToARGBRow = J422ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      J422ToARGBRow = J422ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_J422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    J422ToARGBRow = J422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      J422ToARGBRow = J422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_J422TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    J422ToARGBRow = J422ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      J422ToARGBRow = J422ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_J422TOARGBROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    J422ToARGBRow = J422ToARGBRow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    J422ToARGBRow(src_y, src_u, src_v, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert J422 to ARGB.
+LIBYUV_API
+int J422ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*J422ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = J422ToARGBRow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0;
+  }
+#if defined(HAS_J422TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    J422ToARGBRow = J422ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      J422ToARGBRow = J422ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_J422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    J422ToARGBRow = J422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      J422ToARGBRow = J422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_J422TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    J422ToARGBRow = J422ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      J422ToARGBRow = J422ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_J422TOARGBROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    J422ToARGBRow = J422ToARGBRow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    J422ToARGBRow(src_y, src_u, src_v, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
diff --git a/third_party/libyuv/source/convert_from.cc b/third_party/libyuv/source/convert_from.cc
index c1a2f62f0..b743cde26 100644
--- a/third_party/libyuv/source/convert_from.cc
+++ b/third_party/libyuv/source/convert_from.cc
@@ -13,7 +13,6 @@
 #include "libyuv/basic_types.h"
 #include "libyuv/convert.h"  // For I420Copy
 #include "libyuv/cpu_id.h"
-#include "libyuv/format_conversion.h"
 #include "libyuv/planar_functions.h"
 #include "libyuv/rotate.h"
 #include "libyuv/scale.h"  // For ScalePlane()
@@ -174,14 +173,15 @@ int I422ToYUY2(const uint8* src_y, int src_stride_y,
     src_stride_y = src_stride_u = src_stride_v = dst_stride_yuy2 = 0;
   }
 #if defined(HAS_I422TOYUY2ROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     I422ToYUY2Row = I422ToYUY2Row_Any_SSE2;
     if (IS_ALIGNED(width, 16)) {
       I422ToYUY2Row = I422ToYUY2Row_SSE2;
     }
   }
-#elif defined(HAS_I422TOYUY2ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+#endif
+#if defined(HAS_I422TOYUY2ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     I422ToYUY2Row = I422ToYUY2Row_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
       I422ToYUY2Row = I422ToYUY2Row_NEON;
@@ -220,14 +220,15 @@ int I420ToYUY2(const uint8* src_y, int src_stride_y,
     dst_stride_yuy2 = -dst_stride_yuy2;
   }
 #if defined(HAS_I422TOYUY2ROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     I422ToYUY2Row = I422ToYUY2Row_Any_SSE2;
     if (IS_ALIGNED(width, 16)) {
       I422ToYUY2Row = I422ToYUY2Row_SSE2;
     }
   }
-#elif defined(HAS_I422TOYUY2ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+#endif
+#if defined(HAS_I422TOYUY2ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     I422ToYUY2Row = I422ToYUY2Row_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
       I422ToYUY2Row = I422ToYUY2Row_NEON;
@@ -280,14 +281,15 @@ int I422ToUYVY(const uint8* src_y, int src_stride_y,
     src_stride_y = src_stride_u = src_stride_v = dst_stride_uyvy = 0;
   }
 #if defined(HAS_I422TOUYVYROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     I422ToUYVYRow = I422ToUYVYRow_Any_SSE2;
     if (IS_ALIGNED(width, 16)) {
       I422ToUYVYRow = I422ToUYVYRow_SSE2;
     }
   }
-#elif defined(HAS_I422TOUYVYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+#endif
+#if defined(HAS_I422TOUYVYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     I422ToUYVYRow = I422ToUYVYRow_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
       I422ToUYVYRow = I422ToUYVYRow_NEON;
@@ -326,14 +328,15 @@ int I420ToUYVY(const uint8* src_y, int src_stride_y,
     dst_stride_uyvy = -dst_stride_uyvy;
   }
 #if defined(HAS_I422TOUYVYROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     I422ToUYVYRow = I422ToUYVYRow_Any_SSE2;
     if (IS_ALIGNED(width, 16)) {
       I422ToUYVYRow = I422ToUYVYRow_SSE2;
     }
   }
-#elif defined(HAS_I422TOUYVYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+#endif
+#if defined(HAS_I422TOUYVYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     I422ToUYVYRow = I422ToUYVYRow_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
       I422ToUYVYRow = I422ToUYVYRow_NEON;
@@ -397,20 +400,15 @@ int I420ToNV12(const uint8* src_y, int src_stride_y,
     src_stride_u = src_stride_v = dst_stride_uv = 0;
   }
 #if defined(HAS_MERGEUVROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && halfwidth >= 16) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     MergeUVRow_ = MergeUVRow_Any_SSE2;
     if (IS_ALIGNED(halfwidth, 16)) {
-      MergeUVRow_ = MergeUVRow_Unaligned_SSE2;
-      if (IS_ALIGNED(src_u, 16) && IS_ALIGNED(src_stride_u, 16) &&
-          IS_ALIGNED(src_v, 16) && IS_ALIGNED(src_stride_v, 16) &&
-          IS_ALIGNED(dst_uv, 16) && IS_ALIGNED(dst_stride_uv, 16)) {
-        MergeUVRow_ = MergeUVRow_SSE2;
-      }
+      MergeUVRow_ = MergeUVRow_SSE2;
     }
   }
 #endif
 #if defined(HAS_MERGEUVROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && halfwidth >= 32) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     MergeUVRow_ = MergeUVRow_Any_AVX2;
     if (IS_ALIGNED(halfwidth, 32)) {
       MergeUVRow_ = MergeUVRow_AVX2;
@@ -418,7 +416,7 @@ int I420ToNV12(const uint8* src_y, int src_stride_y,
   }
 #endif
 #if defined(HAS_MERGEUVROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && halfwidth >= 16) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     MergeUVRow_ = MergeUVRow_Any_NEON;
     if (IS_ALIGNED(halfwidth, 16)) {
       MergeUVRow_ = MergeUVRow_NEON;
@@ -476,18 +474,15 @@ int I420ToARGB(const uint8* src_y, int src_stride_y,
     dst_stride_argb = -dst_stride_argb;
   }
 #if defined(HAS_I422TOARGBROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
-      I422ToARGBRow = I422ToARGBRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
-        I422ToARGBRow = I422ToARGBRow_SSSE3;
-      }
+      I422ToARGBRow = I422ToARGBRow_SSSE3;
     }
   }
 #endif
 #if defined(HAS_I422TOARGBROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && width >= 16) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     I422ToARGBRow = I422ToARGBRow_Any_AVX2;
     if (IS_ALIGNED(width, 16)) {
       I422ToARGBRow = I422ToARGBRow_AVX2;
@@ -495,7 +490,7 @@ int I420ToARGB(const uint8* src_y, int src_stride_y,
   }
 #endif
 #if defined(HAS_I422TOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     I422ToARGBRow = I422ToARGBRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       I422ToARGBRow = I422ToARGBRow_NEON;
@@ -548,23 +543,30 @@ int I420ToBGRA(const uint8* src_y, int src_stride_y,
     dst_stride_bgra = -dst_stride_bgra;
   }
 #if defined(HAS_I422TOBGRAROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     I422ToBGRARow = I422ToBGRARow_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
-      I422ToBGRARow = I422ToBGRARow_Unaligned_SSSE3;
-      if (IS_ALIGNED(dst_bgra, 16) && IS_ALIGNED(dst_stride_bgra, 16)) {
-        I422ToBGRARow = I422ToBGRARow_SSSE3;
-      }
+      I422ToBGRARow = I422ToBGRARow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOBGRAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToBGRARow = I422ToBGRARow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToBGRARow = I422ToBGRARow_AVX2;
     }
   }
-#elif defined(HAS_I422TOBGRAROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_I422TOBGRAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     I422ToBGRARow = I422ToBGRARow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       I422ToBGRARow = I422ToBGRARow_NEON;
     }
   }
-#elif defined(HAS_I422TOBGRAROW_MIPS_DSPR2)
+#endif
+#if defined(HAS_I422TOBGRAROW_MIPS_DSPR2)
   if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
       IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
       IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
@@ -610,17 +612,23 @@ int I420ToABGR(const uint8* src_y, int src_stride_y,
     dst_stride_abgr = -dst_stride_abgr;
   }
 #if defined(HAS_I422TOABGRROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     I422ToABGRRow = I422ToABGRRow_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
-      I422ToABGRRow = I422ToABGRRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(dst_abgr, 16) && IS_ALIGNED(dst_stride_abgr, 16)) {
-        I422ToABGRRow = I422ToABGRRow_SSSE3;
-      }
+      I422ToABGRRow = I422ToABGRRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOABGRROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToABGRRow = I422ToABGRRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToABGRRow = I422ToABGRRow_AVX2;
     }
   }
-#elif defined(HAS_I422TOABGRROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_I422TOABGRROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     I422ToABGRRow = I422ToABGRRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       I422ToABGRRow = I422ToABGRRow_NEON;
@@ -664,17 +672,23 @@ int I420ToRGBA(const uint8* src_y, int src_stride_y,
     dst_stride_rgba = -dst_stride_rgba;
   }
 #if defined(HAS_I422TORGBAROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     I422ToRGBARow = I422ToRGBARow_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
-      I422ToRGBARow = I422ToRGBARow_Unaligned_SSSE3;
-      if (IS_ALIGNED(dst_rgba, 16) && IS_ALIGNED(dst_stride_rgba, 16)) {
-        I422ToRGBARow = I422ToRGBARow_SSSE3;
-      }
+      I422ToRGBARow = I422ToRGBARow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToRGBARow = I422ToRGBARow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGBARow = I422ToRGBARow_AVX2;
     }
   }
-#elif defined(HAS_I422TORGBAROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_I422TORGBAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     I422ToRGBARow = I422ToRGBARow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       I422ToRGBARow = I422ToRGBARow_NEON;
@@ -718,14 +732,15 @@ int I420ToRGB24(const uint8* src_y, int src_stride_y,
     dst_stride_rgb24 = -dst_stride_rgb24;
   }
 #if defined(HAS_I422TORGB24ROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     I422ToRGB24Row = I422ToRGB24Row_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
       I422ToRGB24Row = I422ToRGB24Row_SSSE3;
     }
   }
-#elif defined(HAS_I422TORGB24ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_I422TORGB24ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     I422ToRGB24Row = I422ToRGB24Row_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       I422ToRGB24Row = I422ToRGB24Row_NEON;
@@ -769,14 +784,15 @@ int I420ToRAW(const uint8* src_y, int src_stride_y,
     dst_stride_raw = -dst_stride_raw;
   }
 #if defined(HAS_I422TORAWROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     I422ToRAWRow = I422ToRAWRow_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
       I422ToRAWRow = I422ToRAWRow_SSSE3;
     }
   }
-#elif defined(HAS_I422TORAWROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_I422TORAWROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     I422ToRAWRow = I422ToRAWRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       I422ToRAWRow = I422ToRAWRow_NEON;
@@ -820,14 +836,23 @@ int I420ToARGB1555(const uint8* src_y, int src_stride_y,
     dst_stride_argb1555 = -dst_stride_argb1555;
   }
 #if defined(HAS_I422TOARGB1555ROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     I422ToARGB1555Row = I422ToARGB1555Row_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
       I422ToARGB1555Row = I422ToARGB1555Row_SSSE3;
     }
   }
-#elif defined(HAS_I422TOARGB1555ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_I422TOARGB1555ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToARGB1555Row = I422ToARGB1555Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGB1555Row = I422ToARGB1555Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB1555ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     I422ToARGB1555Row = I422ToARGB1555Row_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       I422ToARGB1555Row = I422ToARGB1555Row_NEON;
@@ -872,14 +897,23 @@ int I420ToARGB4444(const uint8* src_y, int src_stride_y,
     dst_stride_argb4444 = -dst_stride_argb4444;
   }
 #if defined(HAS_I422TOARGB4444ROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     I422ToARGB4444Row = I422ToARGB4444Row_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
       I422ToARGB4444Row = I422ToARGB4444Row_SSSE3;
     }
   }
-#elif defined(HAS_I422TOARGB4444ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_I422TOARGB4444ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToARGB4444Row = I422ToARGB4444Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGB4444Row = I422ToARGB4444Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB4444ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     I422ToARGB4444Row = I422ToARGB4444Row_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       I422ToARGB4444Row = I422ToARGB4444Row_NEON;
@@ -923,14 +957,23 @@ int I420ToRGB565(const uint8* src_y, int src_stride_y,
     dst_stride_rgb565 = -dst_stride_rgb565;
   }
 #if defined(HAS_I422TORGB565ROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     I422ToRGB565Row = I422ToRGB565Row_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
       I422ToRGB565Row = I422ToRGB565Row_SSSE3;
     }
   }
-#elif defined(HAS_I422TORGB565ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_I422TORGB565ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGB565Row = I422ToRGB565Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     I422ToRGB565Row = I422ToRGB565Row_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       I422ToRGB565Row = I422ToRGB565Row_NEON;
@@ -1054,38 +1097,6 @@ int ConvertFromI420(const uint8* y, int y_stride,
                      dst_sample_stride ? dst_sample_stride : width * 4,
                      width, height);
       break;
-    case FOURCC_BGGR:
-      r = I420ToBayerBGGR(y, y_stride,
-                          u, u_stride,
-                          v, v_stride,
-                          dst_sample,
-                          dst_sample_stride ? dst_sample_stride : width,
-                          width, height);
-      break;
-    case FOURCC_GBRG:
-      r = I420ToBayerGBRG(y, y_stride,
-                          u, u_stride,
-                          v, v_stride,
-                          dst_sample,
-                          dst_sample_stride ? dst_sample_stride : width,
-                          width, height);
-      break;
-    case FOURCC_GRBG:
-      r = I420ToBayerGRBG(y, y_stride,
-                          u, u_stride,
-                          v, v_stride,
-                          dst_sample,
-                          dst_sample_stride ? dst_sample_stride : width,
-                          width, height);
-      break;
-    case FOURCC_RGGB:
-      r = I420ToBayerRGGB(y, y_stride,
-                          u, u_stride,
-                          v, v_stride,
-                          dst_sample,
-                          dst_sample_stride ? dst_sample_stride : width,
-                          width, height);
-      break;
     case FOURCC_I400:
       r = I400Copy(y, y_stride,
                    dst_sample,
@@ -1116,7 +1127,7 @@ int ConvertFromI420(const uint8* y, int y_stride,
                      width, height);
       break;
     }
-    // TODO(fbarchard): Add M420 and Q420.
+    // TODO(fbarchard): Add M420.
     // Triplanar formats
     // TODO(fbarchard): halfstride instead of halfwidth
     case FOURCC_I420:
diff --git a/third_party/libyuv/source/convert_from_argb.cc b/third_party/libyuv/source/convert_from_argb.cc
index de461ddb0..dc2186a6a 100644
--- a/third_party/libyuv/source/convert_from_argb.cc
+++ b/third_party/libyuv/source/convert_from_argb.cc
@@ -12,7 +12,6 @@
 
 #include "libyuv/basic_types.h"
 #include "libyuv/cpu_id.h"
-#include "libyuv/format_conversion.h"
 #include "libyuv/planar_functions.h"
 #include "libyuv/row.h"
 
@@ -51,17 +50,15 @@ int ARGBToI444(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0;
   }
 #if defined(HAS_ARGBTOUV444ROW_SSSE3)
-    if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    if (TestCpuFlag(kCpuHasSSSE3)) {
       ARGBToUV444Row = ARGBToUV444Row_Any_SSSE3;
       if (IS_ALIGNED(width, 16)) {
-        ARGBToUV444Row = ARGBToUV444Row_Unaligned_SSSE3;
-        if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
-          ARGBToUV444Row = ARGBToUV444Row_SSSE3;
-        }
+        ARGBToUV444Row = ARGBToUV444Row_SSSE3;
       }
   }
-#elif defined(HAS_ARGBTOUV444ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_ARGBTOUV444ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToUV444Row = ARGBToUV444Row_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGBToUV444Row = ARGBToUV444Row_NEON;
@@ -69,19 +66,16 @@ int ARGBToI444(const uint8* src_argb, int src_stride_argb,
   }
 #endif
 #if defined(HAS_ARGBTOYROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
-          IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-        ARGBToYRow = ARGBToYRow_SSSE3;
-      }
+      ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 
-#elif defined(HAS_ARGBTOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGBToYRow = ARGBToYRow_NEON;
@@ -130,17 +124,15 @@ int ARGBToI422(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0;
   }
 #if defined(HAS_ARGBTOUV422ROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToUV422Row = ARGBToUV422Row_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUV422Row = ARGBToUV422Row_Unaligned_SSSE3;
-      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
-        ARGBToUV422Row = ARGBToUV422Row_SSSE3;
-      }
+      ARGBToUV422Row = ARGBToUV422Row_SSSE3;
     }
   }
-#elif defined(HAS_ARGBTOUV422ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+#endif
+#if defined(HAS_ARGBTOUV422ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToUV422Row = ARGBToUV422Row_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
       ARGBToUV422Row = ARGBToUV422Row_NEON;
@@ -149,18 +141,15 @@ int ARGBToI422(const uint8* src_argb, int src_stride_argb,
 #endif
 
 #if defined(HAS_ARGBTOYROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
-          IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-        ARGBToYRow = ARGBToYRow_SSSE3;
-      }
+      ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
-#elif defined(HAS_ARGBTOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGBToYRow = ARGBToYRow_NEON;
@@ -209,19 +198,15 @@ int ARGBToI411(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0;
   }
 #if defined(HAS_ARGBTOYROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
-          IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-        ARGBToYRow = ARGBToYRow_SSSE3;
-      }
+      ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
 #if defined(HAS_ARGBTOYROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     ARGBToYRow = ARGBToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
       ARGBToYRow = ARGBToYRow_AVX2;
@@ -229,7 +214,7 @@ int ARGBToI411(const uint8* src_argb, int src_stride_argb,
   }
 #endif
 #if defined(HAS_ARGBTOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGBToYRow = ARGBToYRow_NEON;
@@ -237,7 +222,7 @@ int ARGBToI411(const uint8* src_argb, int src_stride_argb,
   }
 #endif
 #if defined(HAS_ARGBTOUV411ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 32) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToUV411Row = ARGBToUV411Row_Any_NEON;
     if (IS_ALIGNED(width, 32)) {
       ARGBToUV411Row = ARGBToUV411Row_NEON;
@@ -281,22 +266,17 @@ int ARGBToNV12(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = -src_stride_argb;
   }
 #if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_Unaligned_SSSE3;
-      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
-        ARGBToUVRow = ARGBToUVRow_SSSE3;
-        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-          ARGBToYRow = ARGBToYRow_SSSE3;
-        }
-      }
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+      ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
-#elif defined(HAS_ARGBTOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGBToYRow = ARGBToYRow_NEON;
@@ -304,7 +284,7 @@ int ARGBToNV12(const uint8* src_argb, int src_stride_argb,
   }
 #endif
 #if defined(HAS_ARGBTOUVROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToUVRow = ARGBToUVRow_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
       ARGBToUVRow = ARGBToUVRow_NEON;
@@ -312,18 +292,15 @@ int ARGBToNV12(const uint8* src_argb, int src_stride_argb,
   }
 #endif
 #if defined(HAS_MERGEUVROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && halfwidth >= 16) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     MergeUVRow_ = MergeUVRow_Any_SSE2;
     if (IS_ALIGNED(halfwidth, 16)) {
-      MergeUVRow_ = MergeUVRow_Unaligned_SSE2;
-      if (IS_ALIGNED(dst_uv, 16) && IS_ALIGNED(dst_stride_uv, 16)) {
-        MergeUVRow_ = MergeUVRow_SSE2;
-      }
+      MergeUVRow_ = MergeUVRow_SSE2;
     }
   }
 #endif
 #if defined(HAS_MERGEUVROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && halfwidth >= 32) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     MergeUVRow_ = MergeUVRow_Any_AVX2;
     if (IS_ALIGNED(halfwidth, 32)) {
       MergeUVRow_ = MergeUVRow_AVX2;
@@ -331,7 +308,7 @@ int ARGBToNV12(const uint8* src_argb, int src_stride_argb,
   }
 #endif
 #if defined(HAS_MERGEUVROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && halfwidth >= 16) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     MergeUVRow_ = MergeUVRow_Any_NEON;
     if (IS_ALIGNED(halfwidth, 16)) {
       MergeUVRow_ = MergeUVRow_NEON;
@@ -388,22 +365,17 @@ int ARGBToNV21(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = -src_stride_argb;
   }
 #if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_Unaligned_SSSE3;
-      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
-        ARGBToUVRow = ARGBToUVRow_SSSE3;
-        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-          ARGBToYRow = ARGBToYRow_SSSE3;
-        }
-      }
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+      ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
-#elif defined(HAS_ARGBTOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGBToYRow = ARGBToYRow_NEON;
@@ -411,7 +383,7 @@ int ARGBToNV21(const uint8* src_argb, int src_stride_argb,
   }
 #endif
 #if defined(HAS_ARGBTOUVROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToUVRow = ARGBToUVRow_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
       ARGBToUVRow = ARGBToUVRow_NEON;
@@ -419,18 +391,15 @@ int ARGBToNV21(const uint8* src_argb, int src_stride_argb,
   }
 #endif
 #if defined(HAS_MERGEUVROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && halfwidth >= 16) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     MergeUVRow_ = MergeUVRow_Any_SSE2;
     if (IS_ALIGNED(halfwidth, 16)) {
-      MergeUVRow_ = MergeUVRow_Unaligned_SSE2;
-      if (IS_ALIGNED(dst_uv, 16) && IS_ALIGNED(dst_stride_uv, 16)) {
-        MergeUVRow_ = MergeUVRow_SSE2;
-      }
+      MergeUVRow_ = MergeUVRow_SSE2;
     }
   }
 #endif
 #if defined(HAS_MERGEUVROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && halfwidth >= 32) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     MergeUVRow_ = MergeUVRow_Any_AVX2;
     if (IS_ALIGNED(halfwidth, 32)) {
       MergeUVRow_ = MergeUVRow_AVX2;
@@ -438,7 +407,7 @@ int ARGBToNV21(const uint8* src_argb, int src_stride_argb,
   }
 #endif
 #if defined(HAS_MERGEUVROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && halfwidth >= 16) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     MergeUVRow_ = MergeUVRow_Any_NEON;
     if (IS_ALIGNED(halfwidth, 16)) {
       MergeUVRow_ = MergeUVRow_NEON;
@@ -500,17 +469,15 @@ int ARGBToYUY2(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = dst_stride_yuy2 = 0;
   }
 #if defined(HAS_ARGBTOUV422ROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToUV422Row = ARGBToUV422Row_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUV422Row = ARGBToUV422Row_Unaligned_SSSE3;
-      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
-        ARGBToUV422Row = ARGBToUV422Row_SSSE3;
-      }
+      ARGBToUV422Row = ARGBToUV422Row_SSSE3;
     }
   }
-#elif defined(HAS_ARGBTOUV422ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+#endif
+#if defined(HAS_ARGBTOUV422ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToUV422Row = ARGBToUV422Row_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
       ARGBToUV422Row = ARGBToUV422Row_NEON;
@@ -518,17 +485,15 @@ int ARGBToYUY2(const uint8* src_argb, int src_stride_argb,
   }
 #endif
 #if defined(HAS_ARGBTOYROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
-        ARGBToYRow = ARGBToYRow_SSSE3;
-      }
+      ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
-#elif defined(HAS_ARGBTOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGBToYRow = ARGBToYRow_NEON;
@@ -537,14 +502,15 @@ int ARGBToYUY2(const uint8* src_argb, int src_stride_argb,
 #endif
 
 #if defined(HAS_I422TOYUY2ROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     I422ToYUY2Row = I422ToYUY2Row_Any_SSE2;
     if (IS_ALIGNED(width, 16)) {
       I422ToYUY2Row = I422ToYUY2Row_SSE2;
     }
   }
-#elif defined(HAS_I422TOYUY2ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+#endif
+#if defined(HAS_I422TOYUY2ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     I422ToYUY2Row = I422ToYUY2Row_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
       I422ToYUY2Row = I422ToYUY2Row_NEON;
@@ -602,17 +568,15 @@ int ARGBToUYVY(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = dst_stride_uyvy = 0;
   }
 #if defined(HAS_ARGBTOUV422ROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToUV422Row = ARGBToUV422Row_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUV422Row = ARGBToUV422Row_Unaligned_SSSE3;
-      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
-        ARGBToUV422Row = ARGBToUV422Row_SSSE3;
-      }
+      ARGBToUV422Row = ARGBToUV422Row_SSSE3;
     }
   }
-#elif defined(HAS_ARGBTOUV422ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+#endif
+#if defined(HAS_ARGBTOUV422ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToUV422Row = ARGBToUV422Row_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
       ARGBToUV422Row = ARGBToUV422Row_NEON;
@@ -620,17 +584,15 @@ int ARGBToUYVY(const uint8* src_argb, int src_stride_argb,
   }
 #endif
 #if defined(HAS_ARGBTOYROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
-        ARGBToYRow = ARGBToYRow_SSSE3;
-      }
+      ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
-#elif defined(HAS_ARGBTOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGBToYRow = ARGBToYRow_NEON;
@@ -639,14 +601,15 @@ int ARGBToUYVY(const uint8* src_argb, int src_stride_argb,
 #endif
 
 #if defined(HAS_I422TOUYVYROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     I422ToUYVYRow = I422ToUYVYRow_Any_SSE2;
     if (IS_ALIGNED(width, 16)) {
       I422ToUYVYRow = I422ToUYVYRow_SSE2;
     }
   }
-#elif defined(HAS_I422TOUYVYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+#endif
+#if defined(HAS_I422TOUYVYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     I422ToUYVYRow = I422ToUYVYRow_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
       I422ToUYVYRow = I422ToUYVYRow_NEON;
@@ -697,19 +660,15 @@ int ARGBToI400(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = dst_stride_y = 0;
   }
 #if defined(HAS_ARGBTOYROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
-          IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-        ARGBToYRow = ARGBToYRow_SSSE3;
-      }
+      ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
 #if defined(HAS_ARGBTOYROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     ARGBToYRow = ARGBToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
       ARGBToYRow = ARGBToYRow_AVX2;
@@ -717,7 +676,7 @@ int ARGBToI400(const uint8* src_argb, int src_stride_argb,
   }
 #endif
 #if defined(HAS_ARGBTOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGBToYRow = ARGBToYRow_NEON;
@@ -773,14 +732,15 @@ int ARGBToRGB24(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = dst_stride_rgb24 = 0;
   }
 #if defined(HAS_ARGBTORGB24ROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToRGB24Row = ARGBToRGB24Row_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
       ARGBToRGB24Row = ARGBToRGB24Row_SSSE3;
     }
   }
-#elif defined(HAS_ARGBTORGB24ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_ARGBTORGB24ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToRGB24Row = ARGBToRGB24Row_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGBToRGB24Row = ARGBToRGB24Row_NEON;
@@ -820,14 +780,15 @@ int ARGBToRAW(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = dst_stride_raw = 0;
   }
 #if defined(HAS_ARGBTORAWROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToRAWRow = ARGBToRAWRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
       ARGBToRAWRow = ARGBToRAWRow_SSSE3;
     }
   }
-#elif defined(HAS_ARGBTORAWROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_ARGBTORAWROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToRAWRow = ARGBToRAWRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGBToRAWRow = ARGBToRAWRow_NEON;
@@ -843,6 +804,46 @@ int ARGBToRAW(const uint8* src_argb, int src_stride_argb,
   return 0;
 }
 
+static const uint8 kDither8x8[64] = {
+  0, 128, 32, 160,  8, 136, 40, 168,
+  192, 64, 224, 96, 200, 72, 232, 104,
+  48, 176, 16, 144, 56, 184, 24, 152,
+  240, 112, 208, 80, 248, 120, 216, 88,
+  12, 140, 44, 172,  4, 132, 36, 164,
+  204, 76, 236, 108, 196, 68, 228, 100,
+  60, 188, 28, 156, 52, 180, 20, 148,
+  252, 124, 220, 92, 244, 116, 212, 84,
+};
+
+// Convert ARGB To RGB565 with 8x8 dither matrix (64 bytes).
+LIBYUV_API
+int ARGBToRGB565Dither(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_rgb565, int dst_stride_rgb565,
+                       const uint8* dither8x8, int width, int height) {
+  int y;
+  void (*ARGBToRGB565DitherRow)(const uint8* src_argb, uint8* dst_rgb,
+      const uint8* dither8x8, int pix) = ARGBToRGB565DitherRow_C;
+  if (!src_argb || !dst_rgb565 || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  if (!dither8x8) {
+    dither8x8 = kDither8x8;
+
+  }
+  for (y = 0; y < height; ++y) {
+    ARGBToRGB565DitherRow(src_argb, dst_rgb565,
+                          dither8x8 + ((y & 7) << 3), width);
+    src_argb += src_stride_argb;
+    dst_rgb565 += dst_stride_rgb565;
+  }
+  return 0;
+}
+
 // Convert ARGB To RGB565.
 LIBYUV_API
 int ARGBToRGB565(const uint8* src_argb, int src_stride_argb,
@@ -867,15 +868,23 @@ int ARGBToRGB565(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = dst_stride_rgb565 = 0;
   }
 #if defined(HAS_ARGBTORGB565ROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 4 &&
-      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     ARGBToRGB565Row = ARGBToRGB565Row_Any_SSE2;
     if (IS_ALIGNED(width, 4)) {
       ARGBToRGB565Row = ARGBToRGB565Row_SSE2;
     }
   }
-#elif defined(HAS_ARGBTORGB565ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_ARGBTORGB565ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToRGB565Row = ARGBToRGB565Row_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB565Row = ARGBToRGB565Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToRGB565Row = ARGBToRGB565Row_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGBToRGB565Row = ARGBToRGB565Row_NEON;
@@ -915,15 +924,23 @@ int ARGBToARGB1555(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = dst_stride_argb1555 = 0;
   }
 #if defined(HAS_ARGBTOARGB1555ROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 4 &&
-      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     ARGBToARGB1555Row = ARGBToARGB1555Row_Any_SSE2;
     if (IS_ALIGNED(width, 4)) {
       ARGBToARGB1555Row = ARGBToARGB1555Row_SSE2;
     }
   }
-#elif defined(HAS_ARGBTOARGB1555ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_ARGBTOARGB1555ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToARGB1555Row = ARGBToARGB1555Row_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToARGB1555Row = ARGBToARGB1555Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOARGB1555ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToARGB1555Row = ARGBToARGB1555Row_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGBToARGB1555Row = ARGBToARGB1555Row_NEON;
@@ -963,15 +980,23 @@ int ARGBToARGB4444(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = dst_stride_argb4444 = 0;
   }
 #if defined(HAS_ARGBTOARGB4444ROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 4 &&
-      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     ARGBToARGB4444Row = ARGBToARGB4444Row_Any_SSE2;
     if (IS_ALIGNED(width, 4)) {
       ARGBToARGB4444Row = ARGBToARGB4444Row_SSE2;
     }
   }
-#elif defined(HAS_ARGBTOARGB4444ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_ARGBTOARGB4444ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToARGB4444Row = ARGBToARGB4444Row_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToARGB4444Row = ARGBToARGB4444Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOARGB4444ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToARGB4444Row = ARGBToARGB4444Row_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGBToARGB4444Row = ARGBToARGB4444Row_NEON;
@@ -1011,23 +1036,17 @@ int ARGBToJ420(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = -src_stride_argb;
   }
 #if defined(HAS_ARGBTOYJROW_SSSE3) && defined(HAS_ARGBTOUVJROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3;
     ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVJRow = ARGBToUVJRow_Unaligned_SSSE3;
-      ARGBToYJRow = ARGBToYJRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
-        ARGBToUVJRow = ARGBToUVJRow_SSSE3;
-        if (IS_ALIGNED(dst_yj, 16) && IS_ALIGNED(dst_stride_yj, 16)) {
-          ARGBToYJRow = ARGBToYJRow_SSSE3;
-        }
-      }
+      ARGBToUVJRow = ARGBToUVJRow_SSSE3;
+      ARGBToYJRow = ARGBToYJRow_SSSE3;
     }
   }
 #endif
 #if defined(HAS_ARGBTOYJROW_AVX2) && defined(HAS_ARGBTOUVJROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     ARGBToYJRow = ARGBToYJRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
       ARGBToYJRow = ARGBToYJRow_AVX2;
@@ -1035,7 +1054,7 @@ int ARGBToJ420(const uint8* src_argb, int src_stride_argb,
   }
 #endif
 #if defined(HAS_ARGBTOYJROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYJRow = ARGBToYJRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGBToYJRow = ARGBToYJRow_NEON;
@@ -1043,7 +1062,7 @@ int ARGBToJ420(const uint8* src_argb, int src_stride_argb,
   }
 #endif
 #if defined(HAS_ARGBTOUVJROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToUVJRow = ARGBToUVJRow_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
       ARGBToUVJRow = ARGBToUVJRow_NEON;
@@ -1067,6 +1086,80 @@ int ARGBToJ420(const uint8* src_argb, int src_stride_argb,
   return 0;
 }
 
+// ARGB little endian (bgra in memory) to J422
+LIBYUV_API
+int ARGBToJ422(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*ARGBToUVJ422Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+      int pix) = ARGBToUVJ422Row_C;
+  void (*ARGBToYJRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYJRow_C;
+  if (!src_argb || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_y == width &&
+      dst_stride_u * 2 == width &&
+      dst_stride_v * 2 == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_ARGBTOUVJ422ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVJ422Row = ARGBToUVJ422Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVJ422Row = ARGBToUVJ422Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVJ422ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUVJ422Row = ARGBToUVJ422Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVJ422Row = ARGBToUVJ422Row_NEON;
+    }
+  }
+#endif
+
+#if defined(HAS_ARGBTOYJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYJRow = ARGBToYJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYJRow = ARGBToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYJRow = ARGBToYJRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToUVJ422Row(src_argb, dst_u, dst_v, width);
+    ARGBToYJRow(src_argb, dst_y, width);
+    src_argb += src_stride_argb;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
 // Convert ARGB to J400.
 LIBYUV_API
 int ARGBToJ400(const uint8* src_argb, int src_stride_argb,
@@ -1091,19 +1184,15 @@ int ARGBToJ400(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = dst_stride_yj = 0;
   }
 #if defined(HAS_ARGBTOYJROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToYJRow = ARGBToYJRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
-          IS_ALIGNED(dst_yj, 16) && IS_ALIGNED(dst_stride_yj, 16)) {
-        ARGBToYJRow = ARGBToYJRow_SSSE3;
-      }
+      ARGBToYJRow = ARGBToYJRow_SSSE3;
     }
   }
 #endif
 #if defined(HAS_ARGBTOYJROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     ARGBToYJRow = ARGBToYJRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
       ARGBToYJRow = ARGBToYJRow_AVX2;
@@ -1111,7 +1200,7 @@ int ARGBToJ400(const uint8* src_argb, int src_stride_argb,
   }
 #endif
 #if defined(HAS_ARGBTOYJROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYJRow = ARGBToYJRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGBToYJRow = ARGBToYJRow_NEON;
diff --git a/third_party/libyuv/source/convert_to_argb.cc b/third_party/libyuv/source/convert_to_argb.cc
index 1b228a7b4..af829fbd3 100644
--- a/third_party/libyuv/source/convert_to_argb.cc
+++ b/third_party/libyuv/source/convert_to_argb.cc
@@ -11,7 +11,6 @@
 #include "libyuv/convert_argb.h"
 
 #include "libyuv/cpu_id.h"
-#include "libyuv/format_conversion.h"
 #ifdef HAVE_JPEG
 #include "libyuv/mjpeg_decoder.h"
 #endif
@@ -144,36 +143,6 @@ int ConvertToARGB(const uint8* sample, size_t sample_size,
                          crop_argb, argb_stride,
                          crop_width, inv_crop_height);
       break;
-    // TODO(fbarchard): Support cropping Bayer by odd numbers
-    // by adjusting fourcc.
-    case FOURCC_BGGR:
-      src = sample + (src_width * crop_y + crop_x);
-      r = BayerBGGRToARGB(src, src_width,
-                          crop_argb, argb_stride,
-                          crop_width, inv_crop_height);
-      break;
-
-    case FOURCC_GBRG:
-      src = sample + (src_width * crop_y + crop_x);
-      r = BayerGBRGToARGB(src, src_width,
-                          crop_argb, argb_stride,
-                          crop_width, inv_crop_height);
-      break;
-
-    case FOURCC_GRBG:
-      src = sample + (src_width * crop_y + crop_x);
-      r = BayerGRBGToARGB(src, src_width,
-                          crop_argb, argb_stride,
-                          crop_width, inv_crop_height);
-      break;
-
-    case FOURCC_RGGB:
-      src = sample + (src_width * crop_y + crop_x);
-      r = BayerRGGBToARGB(src, src_width,
-                          crop_argb, argb_stride,
-                          crop_width, inv_crop_height);
-      break;
-
     case FOURCC_I400:
       src = sample + src_width * crop_y + crop_x;
       r = I400ToARGB(src, src_width,
@@ -205,15 +174,6 @@ int ConvertToARGB(const uint8* sample, size_t sample_size,
                      crop_argb, argb_stride,
                      crop_width, inv_crop_height);
       break;
-//    case FOURCC_Q420:
-//      src = sample + (src_width + aligned_src_width * 2) * crop_y + crop_x;
-//      src_uv = sample + (src_width + aligned_src_width * 2) * crop_y +
-//               src_width + crop_x * 2;
-//      r = Q420ToARGB(src, src_width * 3,
-//                    src_uv, src_width * 3,
-//                    crop_argb, argb_stride,
-//                    crop_width, inv_crop_height);
-//      break;
     // Triplanar formats
     case FOURCC_I420:
     case FOURCC_YU12:
@@ -241,6 +201,25 @@ int ConvertToARGB(const uint8* sample, size_t sample_size,
                      crop_width, inv_crop_height);
       break;
     }
+
+    case FOURCC_J420: {
+      const uint8* src_y = sample + (src_width * crop_y + crop_x);
+      const uint8* src_u;
+      const uint8* src_v;
+      int halfwidth = (src_width + 1) / 2;
+      int halfheight = (abs_src_height + 1) / 2;
+      src_u = sample + src_width * abs_src_height +
+          (halfwidth * crop_y + crop_x) / 2;
+      src_v = sample + src_width * abs_src_height +
+          halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      r = J420ToARGB(src_y, src_width,
+                     src_u, halfwidth,
+                     src_v, halfwidth,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+
     case FOURCC_I422:
     case FOURCC_YV16: {
       const uint8* src_y = sample + src_width * crop_y + crop_x;
diff --git a/third_party/libyuv/source/convert_to_i420.cc b/third_party/libyuv/source/convert_to_i420.cc
index 7b194fff7..5e75369b5 100644
--- a/third_party/libyuv/source/convert_to_i420.cc
+++ b/third_party/libyuv/source/convert_to_i420.cc
@@ -12,7 +12,6 @@
 
 #include "libyuv/convert.h"
 
-#include "libyuv/format_conversion.h"
 #include "libyuv/video_common.h"
 
 #ifdef __cplusplus
@@ -173,40 +172,6 @@ int ConvertToI420(const uint8* sample,
                      v, v_stride,
                      crop_width, inv_crop_height);
       break;
-    // TODO(fbarchard): Support cropping Bayer by odd numbers
-    // by adjusting fourcc.
-    case FOURCC_BGGR:
-      src = sample + (src_width * crop_y + crop_x);
-      r = BayerBGGRToI420(src, src_width,
-                          y, y_stride,
-                          u, u_stride,
-                          v, v_stride,
-                          crop_width, inv_crop_height);
-      break;
-    case FOURCC_GBRG:
-      src = sample + (src_width * crop_y + crop_x);
-      r = BayerGBRGToI420(src, src_width,
-                          y, y_stride,
-                          u, u_stride,
-                          v, v_stride,
-                          crop_width, inv_crop_height);
-      break;
-    case FOURCC_GRBG:
-      src = sample + (src_width * crop_y + crop_x);
-      r = BayerGRBGToI420(src, src_width,
-                          y, y_stride,
-                          u, u_stride,
-                          v, v_stride,
-                          crop_width, inv_crop_height);
-      break;
-    case FOURCC_RGGB:
-      src = sample + (src_width * crop_y + crop_x);
-      r = BayerRGGBToI420(src, src_width,
-                          y, y_stride,
-                          u, u_stride,
-                          v, v_stride,
-                          crop_width, inv_crop_height);
-      break;
     case FOURCC_I400:
       src = sample + src_width * crop_y + crop_x;
       r = I400ToI420(src, src_width,
@@ -218,7 +183,8 @@ int ConvertToI420(const uint8* sample,
     // Biplanar formats
     case FOURCC_NV12:
       src = sample + (src_width * crop_y + crop_x);
-      src_uv = sample + aligned_src_width * (src_height + crop_y / 2) + crop_x;
+      src_uv = sample + (src_width * src_height) +
+        ((crop_y / 2) * aligned_src_width) + ((crop_x / 2) * 2);
       r = NV12ToI420Rotate(src, src_width,
                            src_uv, aligned_src_width,
                            y, y_stride,
@@ -228,7 +194,8 @@ int ConvertToI420(const uint8* sample,
       break;
     case FOURCC_NV21:
       src = sample + (src_width * crop_y + crop_x);
-      src_uv = sample + aligned_src_width * (src_height + crop_y / 2) + crop_x;
+      src_uv = sample + (src_width * src_height) +
+        ((crop_y / 2) * aligned_src_width) + ((crop_x / 2) * 2);
       // Call NV12 but with u and v parameters swapped.
       r = NV12ToI420Rotate(src, src_width,
                            src_uv, aligned_src_width,
@@ -245,17 +212,6 @@ int ConvertToI420(const uint8* sample,
                      v, v_stride,
                      crop_width, inv_crop_height);
       break;
-    case FOURCC_Q420:
-      src = sample + (src_width + aligned_src_width * 2) * crop_y + crop_x;
-      src_uv = sample + (src_width + aligned_src_width * 2) * crop_y +
-               src_width + crop_x * 2;
-      r = Q420ToI420(src, src_width * 3,
-                    src_uv, src_width * 3,
-                    y, y_stride,
-                    u, u_stride,
-                    v, v_stride,
-                    crop_width, inv_crop_height);
-      break;
     // Triplanar formats
     case FOURCC_I420:
     case FOURCC_YU12:
diff --git a/third_party/libyuv/source/cpu_id.cc b/third_party/libyuv/source/cpu_id.cc
index 8f8a403ee..1efa26525 100644
--- a/third_party/libyuv/source/cpu_id.cc
+++ b/third_party/libyuv/source/cpu_id.cc
@@ -52,7 +52,8 @@ void CpuId(uint32 info_eax, uint32 info_ecx, uint32* cpu_info) {
 #if defined(_MSC_VER) && !defined(__clang__)
 #if (_MSC_FULL_VER >= 160040219)
   __cpuidex((int*)(cpu_info), info_eax, info_ecx);
-#elif defined(_M_IX86)
+#endif
+#if defined(_M_IX86)
   __asm {
     mov        eax, info_eax
     mov        ecx, info_ecx
@@ -98,13 +99,15 @@ int TestOsSaveYmm() {
   uint32 xcr0 = 0u;
 #if defined(_MSC_VER) && (_MSC_FULL_VER >= 160040219)
   xcr0 = (uint32)(_xgetbv(0));  // VS2010 SP1 required.
-#elif defined(_M_IX86) && defined(_MSC_VER)
+#endif
+#if defined(_M_IX86) && defined(_MSC_VER)
   __asm {
     xor        ecx, ecx    // xcr 0
     _asm _emit 0x0f _asm _emit 0x01 _asm _emit 0xd0  // For VS2010 and earlier.
     mov        xcr0, eax
   }
-#elif defined(__i386__) || defined(__x86_64__)
+#endif
+#if defined(__i386__) || defined(__x86_64__)
   asm(".byte 0x0f, 0x01, 0xd0" : "=a" (xcr0) : "c" (0) : "%edx");
 #endif  // defined(_MSC_VER)
   return((xcr0 & 6) == 6);  // Is ymm saved?
@@ -135,6 +138,12 @@ int ArmCpuCaps(const char* cpuinfo_name) {
         fclose(f);
         return kCpuHasNEON;
       }
+      // aarch64 uses asimd for Neon.
+      p = strstr(cpuinfo_line, " asimd");
+      if (p && (p[6] == ' ' || p[6] == '\n')) {
+        fclose(f);
+        return kCpuHasNEON;
+      }
     }
   }
   fclose(f);
@@ -240,7 +249,8 @@ int InitCpuFlags(void) {
   if (TestEnv("LIBYUV_DISABLE_FMA3")) {
     cpu_info_ &= ~kCpuHasFMA3;
   }
-#elif defined(__mips__) && defined(__linux__)
+#endif
+#if defined(__mips__) && defined(__linux__)
   // Linux mips parse text file for dsp detect.
   cpu_info_ = MipsCpuCaps("dsp");  // set kCpuHasMIPS_DSP.
 #if defined(__mips_dspr2)
@@ -257,7 +267,8 @@ int InitCpuFlags(void) {
   if (getenv("LIBYUV_DISABLE_MIPS_DSPR2")) {
     cpu_info_ &= ~kCpuHasMIPS_DSPR2;
   }
-#elif defined(__arm__) || defined(__aarch64__)
+#endif
+#if defined(__arm__) || defined(__aarch64__)
 // gcc -mfpu=neon defines __ARM_NEON__
 // __ARM_NEON__ generates code that requires Neon.  NaCL also requires Neon.
 // For Linux, /proc/cpuinfo can be tested but without that assume Neon.
@@ -266,7 +277,8 @@ int InitCpuFlags(void) {
 // For aarch64(arm64), /proc/cpuinfo's feature is not complete, e.g. no neon
 // flag in it.
 // So for aarch64, neon enabling is hard coded here.
-#elif defined(__aarch64__)
+#endif
+#if defined(__aarch64__)
   cpu_info_ = kCpuHasNEON;
 #else
   // Linux arm parse text file for neon detect.
diff --git a/third_party/libyuv/source/format_conversion.cc b/third_party/libyuv/source/format_conversion.cc
deleted file mode 100644
index 3c1737153..000000000
--- a/third_party/libyuv/source/format_conversion.cc
+++ /dev/null
@@ -1,554 +0,0 @@
-/*
- *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS. All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-
-#include "libyuv/format_conversion.h"
-
-#include "libyuv/basic_types.h"
-#include "libyuv/cpu_id.h"
-#include "libyuv/video_common.h"
-#include "libyuv/row.h"
-
-#ifdef __cplusplus
-namespace libyuv {
-extern "C" {
-#endif
-
-// generate a selector mask useful for pshufb
-static uint32 GenerateSelector(int select0, int select1) {
-  return (uint32)(select0) |
-         (uint32)((select1 + 4) << 8) |
-         (uint32)((select0 + 8) << 16) |
-         (uint32)((select1 + 12) << 24);
-}
-
-static int MakeSelectors(const int blue_index,
-                         const int green_index,
-                         const int red_index,
-                         uint32 dst_fourcc_bayer,
-                         uint32* index_map) {
-  // Now build a lookup table containing the indices for the four pixels in each
-  // 2x2 Bayer grid.
-  switch (dst_fourcc_bayer) {
-    case FOURCC_BGGR:
-      index_map[0] = GenerateSelector(blue_index, green_index);
-      index_map[1] = GenerateSelector(green_index, red_index);
-      break;
-    case FOURCC_GBRG:
-      index_map[0] = GenerateSelector(green_index, blue_index);
-      index_map[1] = GenerateSelector(red_index, green_index);
-      break;
-    case FOURCC_RGGB:
-      index_map[0] = GenerateSelector(red_index, green_index);
-      index_map[1] = GenerateSelector(green_index, blue_index);
-      break;
-    case FOURCC_GRBG:
-      index_map[0] = GenerateSelector(green_index, red_index);
-      index_map[1] = GenerateSelector(blue_index, green_index);
-      break;
-    default:
-      return -1;  // Bad FourCC
-  }
-  return 0;
-}
-
-// Converts 32 bit ARGB to Bayer RGB formats.
-LIBYUV_API
-int ARGBToBayer(const uint8* src_argb, int src_stride_argb,
-                uint8* dst_bayer, int dst_stride_bayer,
-                int width, int height,
-                uint32 dst_fourcc_bayer) {
-  int y;
-  const int blue_index = 0;  // Offsets for ARGB format
-  const int green_index = 1;
-  const int red_index = 2;
-  uint32 index_map[2];
-  void (*ARGBToBayerRow)(const uint8* src_argb, uint8* dst_bayer,
-                         uint32 selector, int pix) = ARGBToBayerRow_C;
-  if (height < 0) {
-    height = -height;
-    src_argb = src_argb + (height - 1) * src_stride_argb;
-    src_stride_argb = -src_stride_argb;
-  }
-#if defined(HAS_ARGBTOBAYERROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8 &&
-      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
-    ARGBToBayerRow = ARGBToBayerRow_Any_SSSE3;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToBayerRow = ARGBToBayerRow_SSSE3;
-    }
-  }
-#elif defined(HAS_ARGBTOBAYERROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
-    ARGBToBayerRow = ARGBToBayerRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToBayerRow = ARGBToBayerRow_NEON;
-    }
-  }
-#endif
-  if (MakeSelectors(blue_index, green_index, red_index,
-                    dst_fourcc_bayer, index_map)) {
-    return -1;  // Bad FourCC
-  }
-
-  for (y = 0; y < height; ++y) {
-    ARGBToBayerRow(src_argb, dst_bayer, index_map[y & 1], width);
-    src_argb += src_stride_argb;
-    dst_bayer += dst_stride_bayer;
-  }
-  return 0;
-}
-
-#define AVG(a, b) (((a) + (b)) >> 1)
-
-static void BayerRowBG(const uint8* src_bayer0, int src_stride_bayer,
-                       uint8* dst_argb, int pix) {
-  const uint8* src_bayer1 = src_bayer0 + src_stride_bayer;
-  uint8 g = src_bayer0[1];
-  uint8 r = src_bayer1[1];
-  int x;
-  for (x = 0; x < pix - 2; x += 2) {
-    dst_argb[0] = src_bayer0[0];
-    dst_argb[1] = AVG(g, src_bayer0[1]);
-    dst_argb[2] = AVG(r, src_bayer1[1]);
-    dst_argb[3] = 255U;
-    dst_argb[4] = AVG(src_bayer0[0], src_bayer0[2]);
-    dst_argb[5] = src_bayer0[1];
-    dst_argb[6] = src_bayer1[1];
-    dst_argb[7] = 255U;
-    g = src_bayer0[1];
-    r = src_bayer1[1];
-    src_bayer0 += 2;
-    src_bayer1 += 2;
-    dst_argb += 8;
-  }
-  dst_argb[0] = src_bayer0[0];
-  dst_argb[1] = AVG(g, src_bayer0[1]);
-  dst_argb[2] = AVG(r, src_bayer1[1]);
-  dst_argb[3] = 255U;
-  if (!(pix & 1)) {
-    dst_argb[4] = src_bayer0[0];
-    dst_argb[5] = src_bayer0[1];
-    dst_argb[6] = src_bayer1[1];
-    dst_argb[7] = 255U;
-  }
-}
-
-static void BayerRowRG(const uint8* src_bayer0, int src_stride_bayer,
-                       uint8* dst_argb, int pix) {
-  const uint8* src_bayer1 = src_bayer0 + src_stride_bayer;
-  uint8 g = src_bayer0[1];
-  uint8 b = src_bayer1[1];
-  int x;
-  for (x = 0; x < pix - 2; x += 2) {
-    dst_argb[0] = AVG(b, src_bayer1[1]);
-    dst_argb[1] = AVG(g, src_bayer0[1]);
-    dst_argb[2] = src_bayer0[0];
-    dst_argb[3] = 255U;
-    dst_argb[4] = src_bayer1[1];
-    dst_argb[5] = src_bayer0[1];
-    dst_argb[6] = AVG(src_bayer0[0], src_bayer0[2]);
-    dst_argb[7] = 255U;
-    g = src_bayer0[1];
-    b = src_bayer1[1];
-    src_bayer0 += 2;
-    src_bayer1 += 2;
-    dst_argb += 8;
-  }
-  dst_argb[0] = AVG(b, src_bayer1[1]);
-  dst_argb[1] = AVG(g, src_bayer0[1]);
-  dst_argb[2] = src_bayer0[0];
-  dst_argb[3] = 255U;
-  if (!(pix & 1)) {
-    dst_argb[4] = src_bayer1[1];
-    dst_argb[5] = src_bayer0[1];
-    dst_argb[6] = src_bayer0[0];
-    dst_argb[7] = 255U;
-  }
-}
-
-static void BayerRowGB(const uint8* src_bayer0, int src_stride_bayer,
-                       uint8* dst_argb, int pix) {
-  const uint8* src_bayer1 = src_bayer0 + src_stride_bayer;
-  uint8 b = src_bayer0[1];
-  int x;
-  for (x = 0; x < pix - 2; x += 2) {
-    dst_argb[0] = AVG(b, src_bayer0[1]);
-    dst_argb[1] = src_bayer0[0];
-    dst_argb[2] = src_bayer1[0];
-    dst_argb[3] = 255U;
-    dst_argb[4] = src_bayer0[1];
-    dst_argb[5] = AVG(src_bayer0[0], src_bayer0[2]);
-    dst_argb[6] = AVG(src_bayer1[0], src_bayer1[2]);
-    dst_argb[7] = 255U;
-    b = src_bayer0[1];
-    src_bayer0 += 2;
-    src_bayer1 += 2;
-    dst_argb += 8;
-  }
-  dst_argb[0] = AVG(b, src_bayer0[1]);
-  dst_argb[1] = src_bayer0[0];
-  dst_argb[2] = src_bayer1[0];
-  dst_argb[3] = 255U;
-  if (!(pix & 1)) {
-    dst_argb[4] = src_bayer0[1];
-    dst_argb[5] = src_bayer0[0];
-    dst_argb[6] = src_bayer1[0];
-    dst_argb[7] = 255U;
-  }
-}
-
-static void BayerRowGR(const uint8* src_bayer0, int src_stride_bayer,
-                       uint8* dst_argb, int pix) {
-  const uint8* src_bayer1 = src_bayer0 + src_stride_bayer;
-  uint8 r = src_bayer0[1];
-  int x;
-  for (x = 0; x < pix - 2; x += 2) {
-    dst_argb[0] = src_bayer1[0];
-    dst_argb[1] = src_bayer0[0];
-    dst_argb[2] = AVG(r, src_bayer0[1]);
-    dst_argb[3] = 255U;
-    dst_argb[4] = AVG(src_bayer1[0], src_bayer1[2]);
-    dst_argb[5] = AVG(src_bayer0[0], src_bayer0[2]);
-    dst_argb[6] = src_bayer0[1];
-    dst_argb[7] = 255U;
-    r = src_bayer0[1];
-    src_bayer0 += 2;
-    src_bayer1 += 2;
-    dst_argb += 8;
-  }
-  dst_argb[0] = src_bayer1[0];
-  dst_argb[1] = src_bayer0[0];
-  dst_argb[2] = AVG(r, src_bayer0[1]);
-  dst_argb[3] = 255U;
-  if (!(pix & 1)) {
-    dst_argb[4] = src_bayer1[0];
-    dst_argb[5] = src_bayer0[0];
-    dst_argb[6] = src_bayer0[1];
-    dst_argb[7] = 255U;
-  }
-}
-
-// Converts any Bayer RGB format to ARGB.
-LIBYUV_API
-int BayerToARGB(const uint8* src_bayer, int src_stride_bayer,
-                uint8* dst_argb, int dst_stride_argb,
-                int width, int height,
-                uint32 src_fourcc_bayer) {
-  int y;
-  void (*BayerRow0)(const uint8* src_bayer, int src_stride_bayer,
-                    uint8* dst_argb, int pix);
-  void (*BayerRow1)(const uint8* src_bayer, int src_stride_bayer,
-                    uint8* dst_argb, int pix);
-  if (height < 0) {
-    height = -height;
-    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
-    dst_stride_argb = -dst_stride_argb;
-  }
-  switch (src_fourcc_bayer) {
-    case FOURCC_BGGR:
-      BayerRow0 = BayerRowBG;
-      BayerRow1 = BayerRowGR;
-      break;
-    case FOURCC_GBRG:
-      BayerRow0 = BayerRowGB;
-      BayerRow1 = BayerRowRG;
-      break;
-    case FOURCC_GRBG:
-      BayerRow0 = BayerRowGR;
-      BayerRow1 = BayerRowBG;
-      break;
-    case FOURCC_RGGB:
-      BayerRow0 = BayerRowRG;
-      BayerRow1 = BayerRowGB;
-      break;
-    default:
-      return -1;    // Bad FourCC
-  }
-
-  for (y = 0; y < height - 1; y += 2) {
-    BayerRow0(src_bayer, src_stride_bayer, dst_argb, width);
-    BayerRow1(src_bayer + src_stride_bayer, -src_stride_bayer,
-              dst_argb + dst_stride_argb, width);
-    src_bayer += src_stride_bayer * 2;
-    dst_argb += dst_stride_argb * 2;
-  }
-  if (height & 1) {
-    BayerRow0(src_bayer, src_stride_bayer, dst_argb, width);
-  }
-  return 0;
-}
-
-// Converts any Bayer RGB format to ARGB.
-LIBYUV_API
-int BayerToI420(const uint8* src_bayer, int src_stride_bayer,
-                uint8* dst_y, int dst_stride_y,
-                uint8* dst_u, int dst_stride_u,
-                uint8* dst_v, int dst_stride_v,
-                int width, int height,
-                uint32 src_fourcc_bayer) {
-  void (*BayerRow0)(const uint8* src_bayer, int src_stride_bayer,
-                    uint8* dst_argb, int pix);
-  void (*BayerRow1)(const uint8* src_bayer, int src_stride_bayer,
-                    uint8* dst_argb, int pix);
-
-  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
-                      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
-  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
-      ARGBToYRow_C;
-  // Negative height means invert the image.
-  if (height < 0) {
-    int halfheight;
-    height = -height;
-    halfheight = (height + 1) >> 1;
-    dst_y = dst_y + (height - 1) * dst_stride_y;
-    dst_u = dst_u + (halfheight - 1) * dst_stride_u;
-    dst_v = dst_v + (halfheight - 1) * dst_stride_v;
-    dst_stride_y = -dst_stride_y;
-    dst_stride_u = -dst_stride_u;
-    dst_stride_v = -dst_stride_v;
-  }
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
-    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
-    ARGBToYRow = ARGBToYRow_Any_SSSE3;
-    if (IS_ALIGNED(width, 16)) {
-      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
-      ARGBToUVRow = ARGBToUVRow_SSSE3;
-      if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-        ARGBToYRow = ARGBToYRow_SSSE3;
-      }
-    }
-  }
-#elif defined(HAS_ARGBTOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
-    ARGBToYRow = ARGBToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToYRow = ARGBToYRow_NEON;
-    }
-  }
-#endif
-#if defined(HAS_ARGBTOUVROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
-    ARGBToUVRow = ARGBToUVRow_Any_NEON;
-    if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_NEON;
-    }
-  }
-#endif
-
-  switch (src_fourcc_bayer) {
-    case FOURCC_BGGR:
-      BayerRow0 = BayerRowBG;
-      BayerRow1 = BayerRowGR;
-      break;
-    case FOURCC_GBRG:
-      BayerRow0 = BayerRowGB;
-      BayerRow1 = BayerRowRG;
-      break;
-    case FOURCC_GRBG:
-      BayerRow0 = BayerRowGR;
-      BayerRow1 = BayerRowBG;
-      break;
-    case FOURCC_RGGB:
-      BayerRow0 = BayerRowRG;
-      BayerRow1 = BayerRowGB;
-      break;
-    default:
-      return -1;  // Bad FourCC
-  }
-
-  {
-    // Allocate 2 rows of ARGB.
-    const int kRowSize = (width * 4 + 15) & ~15;
-    align_buffer_64(row, kRowSize * 2);
-    int y;
-    for (y = 0; y < height - 1; y += 2) {
-      BayerRow0(src_bayer, src_stride_bayer, row, width);
-      BayerRow1(src_bayer + src_stride_bayer, -src_stride_bayer,
-                row + kRowSize, width);
-      ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
-      ARGBToYRow(row, dst_y, width);
-      ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
-      src_bayer += src_stride_bayer * 2;
-      dst_y += dst_stride_y * 2;
-      dst_u += dst_stride_u;
-      dst_v += dst_stride_v;
-    }
-    if (height & 1) {
-      BayerRow0(src_bayer, src_stride_bayer, row, width);
-      ARGBToUVRow(row, 0, dst_u, dst_v, width);
-      ARGBToYRow(row, dst_y, width);
-    }
-    free_aligned_buffer_64(row);
-  }
-  return 0;
-}
-
-// Convert I420 to Bayer.
-LIBYUV_API
-int I420ToBayer(const uint8* src_y, int src_stride_y,
-                const uint8* src_u, int src_stride_u,
-                const uint8* src_v, int src_stride_v,
-                uint8* dst_bayer, int dst_stride_bayer,
-                int width, int height,
-                uint32 dst_fourcc_bayer) {
-  void (*I422ToARGBRow)(const uint8* y_buf,
-                        const uint8* u_buf,
-                        const uint8* v_buf,
-                        uint8* rgb_buf,
-                        int width) = I422ToARGBRow_C;
-  void (*ARGBToBayerRow)(const uint8* src_argb, uint8* dst_bayer,
-                         uint32 selector, int pix) = ARGBToBayerRow_C;
-  const int blue_index = 0;  // Offsets for ARGB format
-  const int green_index = 1;
-  const int red_index = 2;
-  uint32 index_map[2];
-  // Negative height means invert the image.
-  if (height < 0) {
-    int halfheight;
-    height = -height;
-    halfheight = (height + 1) >> 1;
-    src_y = src_y + (height - 1) * src_stride_y;
-    src_u = src_u + (halfheight - 1) * src_stride_u;
-    src_v = src_v + (halfheight - 1) * src_stride_v;
-    src_stride_y = -src_stride_y;
-    src_stride_u = -src_stride_u;
-    src_stride_v = -src_stride_v;
-  }
-#if defined(HAS_I422TOARGBROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
-    I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToARGBRow = I422ToARGBRow_SSSE3;
-    }
-  }
-#endif
-#if defined(HAS_I422TOARGBROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && width >= 16) {
-    I422ToARGBRow = I422ToARGBRow_Any_AVX2;
-    if (IS_ALIGNED(width, 16)) {
-      I422ToARGBRow = I422ToARGBRow_AVX2;
-    }
-  }
-#endif
-#if defined(HAS_I422TOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
-    I422ToARGBRow = I422ToARGBRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToARGBRow = I422ToARGBRow_NEON;
-    }
-  }
-#endif
-#if defined(HAS_I422TOARGBROW_MIPS_DSPR2)
-  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
-      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
-      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
-      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2)) {
-    I422ToARGBRow = I422ToARGBRow_MIPS_DSPR2;
-  }
-#endif
-
-#if defined(HAS_ARGBTOBAYERROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
-    ARGBToBayerRow = ARGBToBayerRow_Any_SSSE3;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToBayerRow = ARGBToBayerRow_SSSE3;
-    }
-  }
-#elif defined(HAS_ARGBTOBAYERROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
-    ARGBToBayerRow = ARGBToBayerRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToBayerRow = ARGBToBayerRow_NEON;
-    }
-  }
-#endif
-
-  if (MakeSelectors(blue_index, green_index, red_index,
-                    dst_fourcc_bayer, index_map)) {
-    return -1;  // Bad FourCC
-  }
-  {
-    // Allocate a row of ARGB.
-    align_buffer_64(row, width * 4);
-    int y;
-    for (y = 0; y < height; ++y) {
-      I422ToARGBRow(src_y, src_u, src_v, row, width);
-      ARGBToBayerRow(row, dst_bayer, index_map[y & 1], width);
-      dst_bayer += dst_stride_bayer;
-      src_y += src_stride_y;
-      if (y & 1) {
-        src_u += src_stride_u;
-        src_v += src_stride_v;
-      }
-    }
-    free_aligned_buffer_64(row);
-  }
-  return 0;
-}
-
-#define MAKEBAYERFOURCC(BAYER)                                                 \
-LIBYUV_API                                                                     \
-int Bayer##BAYER##ToI420(const uint8* src_bayer, int src_stride_bayer,         \
-                         uint8* dst_y, int dst_stride_y,                       \
-                         uint8* dst_u, int dst_stride_u,                       \
-                         uint8* dst_v, int dst_stride_v,                       \
-                         int width, int height) {                              \
-  return BayerToI420(src_bayer, src_stride_bayer,                              \
-                     dst_y, dst_stride_y,                                      \
-                     dst_u, dst_stride_u,                                      \
-                     dst_v, dst_stride_v,                                      \
-                     width, height,                                            \
-                     FOURCC_##BAYER);                                          \
-}                                                                              \
-                                                                               \
-LIBYUV_API                                                                     \
-int I420ToBayer##BAYER(const uint8* src_y, int src_stride_y,                   \
-                       const uint8* src_u, int src_stride_u,                   \
-                       const uint8* src_v, int src_stride_v,                   \
-                       uint8* dst_bayer, int dst_stride_bayer,                 \
-                       int width, int height) {                                \
-  return I420ToBayer(src_y, src_stride_y,                                      \
-                     src_u, src_stride_u,                                      \
-                     src_v, src_stride_v,                                      \
-                     dst_bayer, dst_stride_bayer,                              \
-                     width, height,                                            \
-                     FOURCC_##BAYER);                                          \
-}                                                                              \
-                                                                               \
-LIBYUV_API                                                                     \
-int ARGBToBayer##BAYER(const uint8* src_argb, int src_stride_argb,             \
-                       uint8* dst_bayer, int dst_stride_bayer,                 \
-                       int width, int height) {                                \
-  return ARGBToBayer(src_argb, src_stride_argb,                                \
-                     dst_bayer, dst_stride_bayer,                              \
-                     width, height,                                            \
-                     FOURCC_##BAYER);                                          \
-}                                                                              \
-                                                                               \
-LIBYUV_API                                                                     \
-int Bayer##BAYER##ToARGB(const uint8* src_bayer, int src_stride_bayer,         \
-                         uint8* dst_argb, int dst_stride_argb,                 \
-                         int width, int height) {                              \
-  return BayerToARGB(src_bayer, src_stride_bayer,                              \
-                     dst_argb, dst_stride_argb,                                \
-                     width, height,                                            \
-                     FOURCC_##BAYER);                                          \
-}
-
-MAKEBAYERFOURCC(BGGR)
-MAKEBAYERFOURCC(GBRG)
-MAKEBAYERFOURCC(GRBG)
-MAKEBAYERFOURCC(RGGB)
-
-#ifdef __cplusplus
-}  // extern "C"
-}  // namespace libyuv
-#endif
diff --git a/third_party/libyuv/source/mjpeg_validate.cc b/third_party/libyuv/source/mjpeg_validate.cc
index 23d22d099..40ce2f787 100644
--- a/third_party/libyuv/source/mjpeg_validate.cc
+++ b/third_party/libyuv/source/mjpeg_validate.cc
@@ -10,15 +10,66 @@
 
 #include "libyuv/mjpeg_decoder.h"
 
+#include <string.h>  // For memchr.
+
 #ifdef __cplusplus
 namespace libyuv {
 extern "C" {
 #endif
 
+// Enable this to try scasb implementation.
+// #define ENABLE_SCASB 1
+
+#ifdef ENABLE_SCASB
+
+// Multiple of 1.
+__declspec(naked) __declspec(align(16))
+const uint8* ScanRow_ERMS(const uint8* src, uint32 val, int count) {
+  __asm {
+    mov        edx, edi
+    mov        edi, [esp + 4]   // src
+    mov        eax, [esp + 8]   // val
+    mov        ecx, [esp + 12]  // count
+    repne scasb
+    jne        sr99
+    mov        eax, edi
+    sub        eax, 1
+    mov        edi, edx
+    ret
+
+  sr99:
+    mov        eax, 0
+    mov        edi, edx
+    ret
+  }
+}
+#endif
+
+// Helper function to scan for EOI marker.
+static LIBYUV_BOOL ScanEOI(const uint8* sample, size_t sample_size) {
+  const uint8* end = sample + sample_size - 1;
+  const uint8* it = sample;
+  for (;;) {
+#ifdef ENABLE_SCASB
+    it = ScanRow_ERMS(it, 0xff, end - it);
+#else
+    it = static_cast<const uint8*>(memchr(it, 0xff, end - it));
+#endif
+    if (it == NULL) {
+      break;
+    }
+    if (it[1] == 0xd9) {
+      return LIBYUV_TRUE;  // Success: Valid jpeg.
+    }
+    ++it;  // Skip over current 0xff.
+  }
+  // ERROR: Invalid jpeg end code not found. Size sample_size
+  return LIBYUV_FALSE;
+}
+
 // Helper function to validate the jpeg appears intact.
-// TODO(fbarchard): Optimize case where SOI is found but EOI is not.
 LIBYUV_BOOL ValidateJpeg(const uint8* sample, size_t sample_size) {
-  size_t i;
+  const size_t kBackSearchSize = 1024;
   if (sample_size < 64) {
     // ERROR: Invalid jpeg size: sample_size
     return LIBYUV_FALSE;
@@ -27,17 +78,20 @@ LIBYUV_BOOL ValidateJpeg(const uint8* sample, size_t sample_size) {
     // ERROR: Invalid jpeg initial start code
     return LIBYUV_FALSE;
   }
-  for (i = sample_size - 2; i > 1;) {
-    if (sample[i] != 0xd9) {
-      if (sample[i] == 0xff && sample[i + 1] == 0xd9) {  // End Of Image
-        return LIBYUV_TRUE;  // Success: Valid jpeg.
-      }
-      --i;
+  // Step over SOI marker.
+  sample += 2;
+  sample_size -= 2;
+
+  // Look for the End Of Image (EOI) marker in the end kilobyte of the buffer.
+  if (sample_size > kBackSearchSize) {
+    if (ScanEOI(sample + sample_size - kBackSearchSize, kBackSearchSize)) {
+      return LIBYUV_TRUE;  // Success: Valid jpeg.
     }
-    --i;
+    // Reduce search size for forward search.
+    sample_size = sample_size - kBackSearchSize + 1;
   }
-  // ERROR: Invalid jpeg end code not found. Size sample_size
-  return LIBYUV_FALSE;
+  return ScanEOI(sample, sample_size);
+
 }
 
 #ifdef __cplusplus
diff --git a/third_party/libyuv/source/planar_functions.cc b/third_party/libyuv/source/planar_functions.cc
index 3857008ca..75ef775dd 100644
--- a/third_party/libyuv/source/planar_functions.cc
+++ b/third_party/libyuv/source/planar_functions.cc
@@ -41,16 +41,14 @@ void CopyPlane(const uint8* src_y, int src_stride_y,
   if (src_y == dst_y && src_stride_y == dst_stride_y) {
     return;
   }
-#if defined(HAS_COPYROW_X86)
-  if (TestCpuFlag(kCpuHasX86) && IS_ALIGNED(width, 4)) {
-    CopyRow = CopyRow_X86;
+#if defined(HAS_COPYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    CopyRow = IS_ALIGNED(width, 32) ? CopyRow_SSE2 : CopyRow_Any_SSE2;
   }
 #endif
-#if defined(HAS_COPYROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 32) &&
-      IS_ALIGNED(src_y, 16) && IS_ALIGNED(src_stride_y, 16) &&
-      IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-    CopyRow = CopyRow_SSE2;
+#if defined(HAS_COPYROW_AVX)
+  if (TestCpuFlag(kCpuHasAVX)) {
+    CopyRow = IS_ALIGNED(width, 64) ? CopyRow_AVX : CopyRow_Any_AVX;
   }
 #endif
 #if defined(HAS_COPYROW_ERMS)
@@ -59,8 +57,8 @@ void CopyPlane(const uint8* src_y, int src_stride_y,
   }
 #endif
 #if defined(HAS_COPYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 32)) {
-    CopyRow = CopyRow_NEON;
+  if (TestCpuFlag(kCpuHasNEON)) {
+    CopyRow = IS_ALIGNED(width, 32) ? CopyRow_NEON : CopyRow_Any_NEON;
   }
 #endif
 #if defined(HAS_COPYROW_MIPS)
@@ -90,15 +88,8 @@ void CopyPlane_16(const uint16* src_y, int src_stride_y,
     height = 1;
     src_stride_y = dst_stride_y = 0;
   }
-#if defined(HAS_COPYROW_16_X86)
-  if (TestCpuFlag(kCpuHasX86) && IS_ALIGNED(width, 4)) {
-    CopyRow = CopyRow_16_X86;
-  }
-#endif
 #if defined(HAS_COPYROW_16_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 32) &&
-      IS_ALIGNED(src_y, 16) && IS_ALIGNED(src_stride_y, 16) &&
-      IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 32)) {
     CopyRow = CopyRow_16_SSE2;
   }
 #endif
@@ -239,25 +230,43 @@ void MirrorPlane(const uint8* src_y, int src_stride_y,
     src_stride_y = -src_stride_y;
   }
 #if defined(HAS_MIRRORROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 16)) {
-    MirrorRow = MirrorRow_NEON;
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MirrorRow = MirrorRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorRow = MirrorRow_NEON;
+    }
   }
 #endif
 #if defined(HAS_MIRRORROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16)) {
-    MirrorRow = MirrorRow_SSE2;
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    MirrorRow = MirrorRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorRow = MirrorRow_SSE2;
+    }
   }
 #endif
 #if defined(HAS_MIRRORROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 16) &&
-      IS_ALIGNED(src_y, 16) && IS_ALIGNED(src_stride_y, 16) &&
-      IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-    MirrorRow = MirrorRow_SSSE3;
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    MirrorRow = MirrorRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorRow = MirrorRow_SSSE3;
+    }
   }
 #endif
 #if defined(HAS_MIRRORROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 32)) {
-    MirrorRow = MirrorRow_AVX2;
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MirrorRow = MirrorRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      MirrorRow = MirrorRow_AVX2;
+    }
+  }
+#endif
+// TODO(fbarchard): Mirror on mips handle unaligned memory.
+#if defined(HAS_MIRRORROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(dst_y, 4) && IS_ALIGNED(dst_stride_y, 4)) {
+    MirrorRow = MirrorRow_MIPS_DSPR2;
   }
 #endif
 
@@ -298,23 +307,17 @@ int YUY2ToI422(const uint8* src_yuy2, int src_stride_yuy2,
     src_stride_yuy2 = dst_stride_y = dst_stride_u = dst_stride_v = 0;
   }
 #if defined(HAS_YUY2TOYROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     YUY2ToUV422Row = YUY2ToUV422Row_Any_SSE2;
     YUY2ToYRow = YUY2ToYRow_Any_SSE2;
     if (IS_ALIGNED(width, 16)) {
-      YUY2ToUV422Row = YUY2ToUV422Row_Unaligned_SSE2;
-      YUY2ToYRow = YUY2ToYRow_Unaligned_SSE2;
-      if (IS_ALIGNED(src_yuy2, 16) && IS_ALIGNED(src_stride_yuy2, 16)) {
-        YUY2ToUV422Row = YUY2ToUV422Row_SSE2;
-        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-          YUY2ToYRow = YUY2ToYRow_SSE2;
-        }
-      }
+      YUY2ToUV422Row = YUY2ToUV422Row_SSE2;
+      YUY2ToYRow = YUY2ToYRow_SSE2;
     }
   }
 #endif
 #if defined(HAS_YUY2TOYROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     YUY2ToUV422Row = YUY2ToUV422Row_Any_AVX2;
     YUY2ToYRow = YUY2ToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
@@ -324,7 +327,7 @@ int YUY2ToI422(const uint8* src_yuy2, int src_stride_yuy2,
   }
 #endif
 #if defined(HAS_YUY2TOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     YUY2ToYRow = YUY2ToYRow_Any_NEON;
     if (width >= 16) {
       YUY2ToUV422Row = YUY2ToUV422Row_Any_NEON;
@@ -376,23 +379,17 @@ int UYVYToI422(const uint8* src_uyvy, int src_stride_uyvy,
     src_stride_uyvy = dst_stride_y = dst_stride_u = dst_stride_v = 0;
   }
 #if defined(HAS_UYVYTOYROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     UYVYToUV422Row = UYVYToUV422Row_Any_SSE2;
     UYVYToYRow = UYVYToYRow_Any_SSE2;
     if (IS_ALIGNED(width, 16)) {
-      UYVYToUV422Row = UYVYToUV422Row_Unaligned_SSE2;
-      UYVYToYRow = UYVYToYRow_Unaligned_SSE2;
-      if (IS_ALIGNED(src_uyvy, 16) && IS_ALIGNED(src_stride_uyvy, 16)) {
-        UYVYToUV422Row = UYVYToUV422Row_SSE2;
-        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-          UYVYToYRow = UYVYToYRow_SSE2;
-        }
-      }
+      UYVYToUV422Row = UYVYToUV422Row_SSE2;
+      UYVYToYRow = UYVYToYRow_SSE2;
     }
   }
 #endif
 #if defined(HAS_UYVYTOYROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     UYVYToUV422Row = UYVYToUV422Row_Any_AVX2;
     UYVYToYRow = UYVYToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
@@ -402,7 +399,7 @@ int UYVYToI422(const uint8* src_uyvy, int src_stride_uyvy,
   }
 #endif
 #if defined(HAS_UYVYTOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     UYVYToYRow = UYVYToYRow_Any_NEON;
     if (width >= 16) {
       UYVYToUV422Row = UYVYToUV422Row_Any_NEON;
@@ -497,22 +494,28 @@ int ARGBMirror(const uint8* src_argb, int src_stride_argb,
     src_argb = src_argb + (height - 1) * src_stride_argb;
     src_stride_argb = -src_stride_argb;
   }
-
-#if defined(HAS_ARGBMIRRORROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 4) &&
-      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
-      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
-    ARGBMirrorRow = ARGBMirrorRow_SSSE3;
+#if defined(HAS_ARGBMIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_NEON;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBMirrorRow = ARGBMirrorRow_NEON;
+    }
   }
 #endif
-#if defined(HAS_ARGBMIRRORROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 8)) {
-    ARGBMirrorRow = ARGBMirrorRow_AVX2;
+#if defined(HAS_ARGBMIRRORROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBMirrorRow = ARGBMirrorRow_SSE2;
+    }
   }
 #endif
-#if defined(HAS_ARGBMIRRORROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 4)) {
-    ARGBMirrorRow = ARGBMirrorRow_NEON;
+#if defined(HAS_ARGBMIRRORROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBMirrorRow = ARGBMirrorRow_AVX2;
+    }
   }
 #endif
 
@@ -614,7 +617,7 @@ int ARGBMultiply(const uint8* src_argb0, int src_stride_argb0,
     src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
   }
 #if defined(HAS_ARGBMULTIPLYROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 4) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     ARGBMultiplyRow = ARGBMultiplyRow_Any_SSE2;
     if (IS_ALIGNED(width, 4)) {
       ARGBMultiplyRow = ARGBMultiplyRow_SSE2;
@@ -622,7 +625,7 @@ int ARGBMultiply(const uint8* src_argb0, int src_stride_argb0,
   }
 #endif
 #if defined(HAS_ARGBMULTIPLYROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     ARGBMultiplyRow = ARGBMultiplyRow_Any_AVX2;
     if (IS_ALIGNED(width, 8)) {
       ARGBMultiplyRow = ARGBMultiplyRow_AVX2;
@@ -630,7 +633,7 @@ int ARGBMultiply(const uint8* src_argb0, int src_stride_argb0,
   }
 #endif
 #if defined(HAS_ARGBMULTIPLYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBMultiplyRow = ARGBMultiplyRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGBMultiplyRow = ARGBMultiplyRow_NEON;
@@ -680,7 +683,7 @@ int ARGBAdd(const uint8* src_argb0, int src_stride_argb0,
   }
 #endif
 #if defined(HAS_ARGBADDROW_SSE2) && !defined(_MSC_VER)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 4) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     ARGBAddRow = ARGBAddRow_Any_SSE2;
     if (IS_ALIGNED(width, 4)) {
       ARGBAddRow = ARGBAddRow_SSE2;
@@ -688,7 +691,7 @@ int ARGBAdd(const uint8* src_argb0, int src_stride_argb0,
   }
 #endif
 #if defined(HAS_ARGBADDROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     ARGBAddRow = ARGBAddRow_Any_AVX2;
     if (IS_ALIGNED(width, 8)) {
       ARGBAddRow = ARGBAddRow_AVX2;
@@ -696,7 +699,7 @@ int ARGBAdd(const uint8* src_argb0, int src_stride_argb0,
   }
 #endif
 #if defined(HAS_ARGBADDROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBAddRow = ARGBAddRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGBAddRow = ARGBAddRow_NEON;
@@ -741,7 +744,7 @@ int ARGBSubtract(const uint8* src_argb0, int src_stride_argb0,
     src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
   }
 #if defined(HAS_ARGBSUBTRACTROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 4) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     ARGBSubtractRow = ARGBSubtractRow_Any_SSE2;
     if (IS_ALIGNED(width, 4)) {
       ARGBSubtractRow = ARGBSubtractRow_SSE2;
@@ -749,7 +752,7 @@ int ARGBSubtract(const uint8* src_argb0, int src_stride_argb0,
   }
 #endif
 #if defined(HAS_ARGBSUBTRACTROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     ARGBSubtractRow = ARGBSubtractRow_Any_AVX2;
     if (IS_ALIGNED(width, 8)) {
       ARGBSubtractRow = ARGBSubtractRow_AVX2;
@@ -757,7 +760,7 @@ int ARGBSubtract(const uint8* src_argb0, int src_stride_argb0,
   }
 #endif
 #if defined(HAS_ARGBSUBTRACTROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBSubtractRow = ARGBSubtractRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGBSubtractRow = ARGBSubtractRow_NEON;
@@ -808,24 +811,31 @@ int I422ToBGRA(const uint8* src_y, int src_stride_y,
     height = 1;
     src_stride_y = src_stride_u = src_stride_v = dst_stride_bgra = 0;
   }
-#if defined(HAS_I422TOBGRAROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    I422ToBGRARow = I422ToBGRARow_Any_NEON;
+#if defined(HAS_I422TOBGRAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToBGRARow = I422ToBGRARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToBGRARow = I422ToBGRARow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOBGRAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToBGRARow = I422ToBGRARow_Any_AVX2;
     if (IS_ALIGNED(width, 16)) {
-      I422ToBGRARow = I422ToBGRARow_NEON;
+      I422ToBGRARow = I422ToBGRARow_AVX2;
     }
   }
-#elif defined(HAS_I422TOBGRAROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
-    I422ToBGRARow = I422ToBGRARow_Any_SSSE3;
+#endif
+#if defined(HAS_I422TOBGRAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToBGRARow = I422ToBGRARow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
-      I422ToBGRARow = I422ToBGRARow_Unaligned_SSSE3;
-      if (IS_ALIGNED(dst_bgra, 16) && IS_ALIGNED(dst_stride_bgra, 16)) {
-        I422ToBGRARow = I422ToBGRARow_SSSE3;
-      }
+      I422ToBGRARow = I422ToBGRARow_NEON;
     }
   }
-#elif defined(HAS_I422TOBGRAROW_MIPS_DSPR2)
+#endif
+#if defined(HAS_I422TOBGRAROW_MIPS_DSPR2)
   if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
       IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
       IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
@@ -879,20 +889,26 @@ int I422ToABGR(const uint8* src_y, int src_stride_y,
     src_stride_y = src_stride_u = src_stride_v = dst_stride_abgr = 0;
   }
 #if defined(HAS_I422TOABGRROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
     I422ToABGRRow = I422ToABGRRow_Any_NEON;
-    if (IS_ALIGNED(width, 16)) {
+    if (IS_ALIGNED(width, 8)) {
       I422ToABGRRow = I422ToABGRRow_NEON;
     }
   }
-#elif defined(HAS_I422TOABGRROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+#endif
+#if defined(HAS_I422TOABGRROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     I422ToABGRRow = I422ToABGRRow_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
-      I422ToABGRRow = I422ToABGRRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(dst_abgr, 16) && IS_ALIGNED(dst_stride_abgr, 16)) {
-        I422ToABGRRow = I422ToABGRRow_SSSE3;
-      }
+      I422ToABGRRow = I422ToABGRRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOABGRROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToABGRRow = I422ToABGRRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToABGRRow = I422ToABGRRow_AVX2;
     }
   }
 #endif
@@ -941,20 +957,26 @@ int I422ToRGBA(const uint8* src_y, int src_stride_y,
     src_stride_y = src_stride_u = src_stride_v = dst_stride_rgba = 0;
   }
 #if defined(HAS_I422TORGBAROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
     I422ToRGBARow = I422ToRGBARow_Any_NEON;
-    if (IS_ALIGNED(width, 16)) {
+    if (IS_ALIGNED(width, 8)) {
       I422ToRGBARow = I422ToRGBARow_NEON;
     }
   }
-#elif defined(HAS_I422TORGBAROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+#endif
+#if defined(HAS_I422TORGBAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     I422ToRGBARow = I422ToRGBARow_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
-      I422ToRGBARow = I422ToRGBARow_Unaligned_SSSE3;
-      if (IS_ALIGNED(dst_rgba, 16) && IS_ALIGNED(dst_stride_rgba, 16)) {
-        I422ToRGBARow = I422ToRGBARow_SSSE3;
-      }
+      I422ToRGBARow = I422ToRGBARow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToRGBARow = I422ToRGBARow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGBARow = I422ToRGBARow_AVX2;
     }
   }
 #endif
@@ -991,14 +1013,23 @@ int NV12ToRGB565(const uint8* src_y, int src_stride_y,
     dst_stride_rgb565 = -dst_stride_rgb565;
   }
 #if defined(HAS_NV12TORGB565ROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     NV12ToRGB565Row = NV12ToRGB565Row_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
       NV12ToRGB565Row = NV12ToRGB565Row_SSSE3;
     }
   }
-#elif defined(HAS_NV12TORGB565ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_NV12TORGB565ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_NV12TORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     NV12ToRGB565Row = NV12ToRGB565Row_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       NV12ToRGB565Row = NV12ToRGB565Row_NEON;
@@ -1039,14 +1070,23 @@ int NV21ToRGB565(const uint8* src_y, int src_stride_y,
     dst_stride_rgb565 = -dst_stride_rgb565;
   }
 #if defined(HAS_NV21TORGB565ROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     NV21ToRGB565Row = NV21ToRGB565Row_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
       NV21ToRGB565Row = NV21ToRGB565Row_SSSE3;
     }
   }
-#elif defined(HAS_NV21TORGB565ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+#endif
+#if defined(HAS_NV21TORGB565ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    NV21ToRGB565Row = NV21ToRGB565Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      NV21ToRGB565Row = NV21ToRGB565Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_NV21TORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
     NV21ToRGB565Row = NV21ToRGB565Row_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       NV21ToRGB565Row = NV21ToRGB565Row_NEON;
@@ -1070,8 +1110,12 @@ void SetPlane(uint8* dst_y, int dst_stride_y,
               int width, int height,
               uint32 value) {
   int y;
-  uint32 v32 = value | (value << 8) | (value << 16) | (value << 24);
-  void (*SetRow)(uint8* dst, uint32 value, int pix) = SetRow_C;
+  void (*SetRow)(uint8* dst, uint8 value, int pix) = SetRow_C;
+  if (height < 0) {
+    height = -height;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_stride_y = -dst_stride_y;
+  }
   // Coalesce rows.
   if (dst_stride_y == width) {
     width *= height;
@@ -1079,21 +1123,30 @@ void SetPlane(uint8* dst_y, int dst_stride_y,
     dst_stride_y = 0;
   }
 #if defined(HAS_SETROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) &&
-      IS_ALIGNED(width, 16) &&
-      IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
-    SetRow = SetRow_NEON;
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SetRow = SetRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      SetRow = SetRow_NEON;
+    }
   }
 #endif
 #if defined(HAS_SETROW_X86)
-  if (TestCpuFlag(kCpuHasX86) && IS_ALIGNED(width, 4)) {
-    SetRow = SetRow_X86;
+  if (TestCpuFlag(kCpuHasX86)) {
+    SetRow = SetRow_Any_X86;
+    if (IS_ALIGNED(width, 4)) {
+      SetRow = SetRow_X86;
+    }
+  }
+#endif
+#if defined(HAS_SETROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    SetRow = SetRow_ERMS;
   }
 #endif
 
   // Set plane
   for (y = 0; y < height; ++y) {
-    SetRow(dst_y, v32, width);
+    SetRow(dst_y, value, width);
     dst_y += dst_stride_y;
   }
 }
@@ -1112,7 +1165,7 @@ int I420Rect(uint8* dst_y, int dst_stride_y,
   uint8* start_u = dst_u + (y / 2) * dst_stride_u + (x / 2);
   uint8* start_v = dst_v + (y / 2) * dst_stride_v + (x / 2);
   if (!dst_y || !dst_u || !dst_v ||
-      width <= 0 || height <= 0 ||
+      width <= 0 || height == 0 ||
       x < 0 || y < 0 ||
       value_y < 0 || value_y > 255 ||
       value_u < 0 || value_u > 255 ||
@@ -1132,11 +1185,18 @@ int ARGBRect(uint8* dst_argb, int dst_stride_argb,
              int dst_x, int dst_y,
              int width, int height,
              uint32 value) {
+  int y;
+  void (*ARGBSetRow)(uint8* dst_argb, uint32 value, int pix) = ARGBSetRow_C;
   if (!dst_argb ||
-      width <= 0 || height <= 0 ||
+      width <= 0 || height == 0 ||
       dst_x < 0 || dst_y < 0) {
     return -1;
   }
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
   dst_argb += dst_y * dst_stride_argb + dst_x * 4;
   // Coalesce rows.
   if (dst_stride_argb == width * 4) {
@@ -1144,20 +1204,26 @@ int ARGBRect(uint8* dst_argb, int dst_stride_argb,
     height = 1;
     dst_stride_argb = 0;
   }
-#if defined(HAS_SETROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 16) &&
-      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
-    ARGBSetRows_NEON(dst_argb, value, width, dst_stride_argb, height);
-    return 0;
+
+#if defined(HAS_ARGBSETROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBSetRow = ARGBSetRow_Any_NEON;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBSetRow = ARGBSetRow_NEON;
+    }
   }
 #endif
-#if defined(HAS_SETROW_X86)
+#if defined(HAS_ARGBSETROW_X86)
   if (TestCpuFlag(kCpuHasX86)) {
-    ARGBSetRows_X86(dst_argb, value, width, dst_stride_argb, height);
-    return 0;
+    ARGBSetRow = ARGBSetRow_X86;
   }
 #endif
-  ARGBSetRows_C(dst_argb, value, width, dst_stride_argb, height);
+
+  // Set plane
+  for (y = 0; y < height; ++y) {
+    ARGBSetRow(dst_argb, value, width);
+    dst_argb += dst_stride_argb;
+  }
   return 0;
 }
 
@@ -1197,9 +1263,7 @@ int ARGBAttenuate(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = dst_stride_argb = 0;
   }
 #if defined(HAS_ARGBATTENUATEROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 4 &&
-      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
-      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     ARGBAttenuateRow = ARGBAttenuateRow_Any_SSE2;
     if (IS_ALIGNED(width, 4)) {
       ARGBAttenuateRow = ARGBAttenuateRow_SSE2;
@@ -1207,7 +1271,7 @@ int ARGBAttenuate(const uint8* src_argb, int src_stride_argb,
   }
 #endif
 #if defined(HAS_ARGBATTENUATEROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 4) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
     if (IS_ALIGNED(width, 4)) {
       ARGBAttenuateRow = ARGBAttenuateRow_SSSE3;
@@ -1215,7 +1279,7 @@ int ARGBAttenuate(const uint8* src_argb, int src_stride_argb,
   }
 #endif
 #if defined(HAS_ARGBATTENUATEROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
     if (IS_ALIGNED(width, 8)) {
       ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
@@ -1223,7 +1287,7 @@ int ARGBAttenuate(const uint8* src_argb, int src_stride_argb,
   }
 #endif
 #if defined(HAS_ARGBATTENUATEROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGBAttenuateRow = ARGBAttenuateRow_NEON;
@@ -1263,7 +1327,7 @@ int ARGBUnattenuate(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = dst_stride_argb = 0;
   }
 #if defined(HAS_ARGBUNATTENUATEROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 4) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     ARGBUnattenuateRow = ARGBUnattenuateRow_Any_SSE2;
     if (IS_ALIGNED(width, 4)) {
       ARGBUnattenuateRow = ARGBUnattenuateRow_SSE2;
@@ -1271,7 +1335,7 @@ int ARGBUnattenuate(const uint8* src_argb, int src_stride_argb,
   }
 #endif
 #if defined(HAS_ARGBUNATTENUATEROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     ARGBUnattenuateRow = ARGBUnattenuateRow_Any_AVX2;
     if (IS_ALIGNED(width, 8)) {
       ARGBUnattenuateRow = ARGBUnattenuateRow_AVX2;
@@ -1312,12 +1376,11 @@ int ARGBGrayTo(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = dst_stride_argb = 0;
   }
 #if defined(HAS_ARGBGRAYROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8) &&
-      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
-      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8)) {
     ARGBGrayRow = ARGBGrayRow_SSSE3;
   }
-#elif defined(HAS_ARGBGRAYROW_NEON)
+#endif
+#if defined(HAS_ARGBGRAYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
     ARGBGrayRow = ARGBGrayRow_NEON;
   }
@@ -1350,11 +1413,11 @@ int ARGBGray(uint8* dst_argb, int dst_stride_argb,
     dst_stride_argb = 0;
   }
 #if defined(HAS_ARGBGRAYROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8) &&
-      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8)) {
     ARGBGrayRow = ARGBGrayRow_SSSE3;
   }
-#elif defined(HAS_ARGBGRAYROW_NEON)
+#endif
+#if defined(HAS_ARGBGRAYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
     ARGBGrayRow = ARGBGrayRow_NEON;
   }
@@ -1383,11 +1446,11 @@ int ARGBSepia(uint8* dst_argb, int dst_stride_argb,
     dst_stride_argb = 0;
   }
 #if defined(HAS_ARGBSEPIAROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8) &&
-      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8)) {
     ARGBSepiaRow = ARGBSepiaRow_SSSE3;
   }
-#elif defined(HAS_ARGBSEPIAROW_NEON)
+#endif
+#if defined(HAS_ARGBSEPIAROW_NEON)
   if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
     ARGBSepiaRow = ARGBSepiaRow_NEON;
   }
@@ -1425,11 +1488,11 @@ int ARGBColorMatrix(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = dst_stride_argb = 0;
   }
 #if defined(HAS_ARGBCOLORMATRIXROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8) &&
-      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8)) {
     ARGBColorMatrixRow = ARGBColorMatrixRow_SSSE3;
   }
-#elif defined(HAS_ARGBCOLORMATRIXROW_NEON)
+#endif
+#if defined(HAS_ARGBCOLORMATRIXROW_NEON)
   if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
     ARGBColorMatrixRow = ARGBColorMatrixRow_NEON;
   }
@@ -1568,11 +1631,11 @@ int ARGBQuantize(uint8* dst_argb, int dst_stride_argb,
     dst_stride_argb = 0;
   }
 #if defined(HAS_ARGBQUANTIZEROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 4) &&
-      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 4)) {
     ARGBQuantizeRow = ARGBQuantizeRow_SSE2;
   }
-#elif defined(HAS_ARGBQUANTIZEROW_NEON)
+#endif
+#if defined(HAS_ARGBQUANTIZEROW_NEON)
   if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
     ARGBQuantizeRow = ARGBQuantizeRow_NEON;
   }
@@ -1743,12 +1806,11 @@ int ARGBShade(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = dst_stride_argb = 0;
   }
 #if defined(HAS_ARGBSHADEROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 4) &&
-      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
-      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 4)) {
     ARGBShadeRow = ARGBShadeRow_SSE2;
   }
-#elif defined(HAS_ARGBSHADEROW_NEON)
+#endif
+#if defined(HAS_ARGBSHADEROW_NEON)
   if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
     ARGBShadeRow = ARGBShadeRow_NEON;
   }
@@ -1790,33 +1852,23 @@ int ARGBInterpolate(const uint8* src_argb0, int src_stride_argb0,
     src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
   }
 #if defined(HAS_INTERPOLATEROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 4) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     InterpolateRow = InterpolateRow_Any_SSE2;
     if (IS_ALIGNED(width, 4)) {
-      InterpolateRow = InterpolateRow_Unaligned_SSE2;
-      if (IS_ALIGNED(src_argb0, 16) && IS_ALIGNED(src_stride_argb0, 16) &&
-          IS_ALIGNED(src_argb1, 16) && IS_ALIGNED(src_stride_argb1, 16) &&
-          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
-        InterpolateRow = InterpolateRow_SSE2;
-      }
+      InterpolateRow = InterpolateRow_SSE2;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 4) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     InterpolateRow = InterpolateRow_Any_SSSE3;
     if (IS_ALIGNED(width, 4)) {
-      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(src_argb0, 16) && IS_ALIGNED(src_stride_argb0, 16) &&
-          IS_ALIGNED(src_argb1, 16) && IS_ALIGNED(src_stride_argb1, 16) &&
-          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
-        InterpolateRow = InterpolateRow_SSSE3;
-      }
+      InterpolateRow = InterpolateRow_SSSE3;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     InterpolateRow = InterpolateRow_Any_AVX2;
     if (IS_ALIGNED(width, 8)) {
       InterpolateRow = InterpolateRow_AVX2;
@@ -1824,19 +1876,19 @@ int ARGBInterpolate(const uint8* src_argb0, int src_stride_argb0,
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 4) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     InterpolateRow = InterpolateRow_Any_NEON;
     if (IS_ALIGNED(width, 4)) {
       InterpolateRow = InterpolateRow_NEON;
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROWS_MIPS_DSPR2)
-  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && width >= 1 &&
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
       IS_ALIGNED(src_argb0, 4) && IS_ALIGNED(src_stride_argb0, 4) &&
       IS_ALIGNED(src_argb1, 4) && IS_ALIGNED(src_stride_argb1, 4) &&
       IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
-    ScaleARGBFilterRows = InterpolateRow_MIPS_DSPR2;
+    InterpolateRow = InterpolateRow_MIPS_DSPR2;
   }
 #endif
 
@@ -1876,7 +1928,7 @@ int ARGBShuffle(const uint8* src_bgra, int src_stride_bgra,
     src_stride_bgra = dst_stride_argb = 0;
   }
 #if defined(HAS_ARGBSHUFFLEROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 4) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     ARGBShuffleRow = ARGBShuffleRow_Any_SSE2;
     if (IS_ALIGNED(width, 4)) {
       ARGBShuffleRow = ARGBShuffleRow_SSE2;
@@ -1884,19 +1936,15 @@ int ARGBShuffle(const uint8* src_bgra, int src_stride_bgra,
   }
 #endif
 #if defined(HAS_ARGBSHUFFLEROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBShuffleRow = ARGBShuffleRow_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
-      ARGBShuffleRow = ARGBShuffleRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(src_bgra, 16) && IS_ALIGNED(src_stride_bgra, 16) &&
-          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
-        ARGBShuffleRow = ARGBShuffleRow_SSSE3;
-      }
+      ARGBShuffleRow = ARGBShuffleRow_SSSE3;
     }
   }
 #endif
 #if defined(HAS_ARGBSHUFFLEROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && width >= 16) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     ARGBShuffleRow = ARGBShuffleRow_Any_AVX2;
     if (IS_ALIGNED(width, 16)) {
       ARGBShuffleRow = ARGBShuffleRow_AVX2;
@@ -1904,7 +1952,7 @@ int ARGBShuffle(const uint8* src_bgra, int src_stride_bgra,
   }
 #endif
 #if defined(HAS_ARGBSHUFFLEROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 4) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBShuffleRow = ARGBShuffleRow_Any_NEON;
     if (IS_ALIGNED(width, 4)) {
       ARGBShuffleRow = ARGBShuffleRow_NEON;
@@ -1947,8 +1995,7 @@ static int ARGBSobelize(const uint8* src_argb, int src_stride_argb,
   }
   // ARGBToBayer used to select G channel from ARGB.
 #if defined(HAS_ARGBTOBAYERGGROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 8 &&
-      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     ARGBToBayerRow = ARGBToBayerGGRow_Any_SSE2;
     if (IS_ALIGNED(width, 8)) {
       ARGBToBayerRow = ARGBToBayerGGRow_SSE2;
@@ -1956,8 +2003,7 @@ static int ARGBSobelize(const uint8* src_argb, int src_stride_argb,
   }
 #endif
 #if defined(HAS_ARGBTOBAYERROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8 &&
-      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToBayerRow = ARGBToBayerRow_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
       ARGBToBayerRow = ARGBToBayerRow_SSSE3;
@@ -1965,7 +2011,7 @@ static int ARGBSobelize(const uint8* src_argb, int src_stride_argb,
   }
 #endif
 #if defined(HAS_ARGBTOBAYERGGROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToBayerRow = ARGBToBayerGGRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGBToBayerRow = ARGBToBayerGGRow_NEON;
@@ -2048,8 +2094,7 @@ int ARGBSobel(const uint8* src_argb, int src_stride_argb,
   void (*SobelRow)(const uint8* src_sobelx, const uint8* src_sobely,
                    uint8* dst_argb, int width) = SobelRow_C;
 #if defined(HAS_SOBELROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16) &&
-      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16)) {
     SobelRow = SobelRow_SSE2;
   }
 #endif
@@ -2070,8 +2115,7 @@ int ARGBSobelToPlane(const uint8* src_argb, int src_stride_argb,
   void (*SobelToPlaneRow)(const uint8* src_sobelx, const uint8* src_sobely,
                           uint8* dst_, int width) = SobelToPlaneRow_C;
 #if defined(HAS_SOBELTOPLANEROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16) &&
-      IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16)) {
     SobelToPlaneRow = SobelToPlaneRow_SSE2;
   }
 #endif
@@ -2093,8 +2137,7 @@ int ARGBSobelXY(const uint8* src_argb, int src_stride_argb,
   void (*SobelXYRow)(const uint8* src_sobelx, const uint8* src_sobely,
                      uint8* dst_argb, int width) = SobelXYRow_C;
 #if defined(HAS_SOBELXYROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16) &&
-      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16)) {
     SobelXYRow = SobelXYRow_SSE2;
   }
 #endif
@@ -2218,10 +2261,7 @@ int ARGBCopyAlpha(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = dst_stride_argb = 0;
   }
 #if defined(HAS_ARGBCOPYALPHAROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) &&
-      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
-      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16) &&
-      IS_ALIGNED(width, 8)) {
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 8)) {
     ARGBCopyAlphaRow = ARGBCopyAlphaRow_SSE2;
   }
 #endif
@@ -2264,10 +2304,7 @@ int ARGBCopyYToAlpha(const uint8* src_y, int src_stride_y,
     src_stride_y = dst_stride_argb = 0;
   }
 #if defined(HAS_ARGBCOPYYTOALPHAROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) &&
-      IS_ALIGNED(src_y, 16) && IS_ALIGNED(src_stride_y, 16) &&
-      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16) &&
-      IS_ALIGNED(width, 8)) {
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 8)) {
     ARGBCopyYToAlphaRow = ARGBCopyYToAlphaRow_SSE2;
   }
 #endif
diff --git a/third_party/libyuv/source/rotate.cc b/third_party/libyuv/source/rotate.cc
index 2ef3228cb..5acaccfd8 100644
--- a/third_party/libyuv/source/rotate.cc
+++ b/third_party/libyuv/source/rotate.cc
@@ -42,11 +42,7 @@ extern "C" {
 #endif
 
 #if !defined(LIBYUV_DISABLE_NEON) && !defined(__native_client__) && \
-    (defined(__ARM_NEON__) || defined(LIBYUV_NEON))
-#define HAS_MIRRORROW_NEON
-void MirrorRow_NEON(const uint8* src, uint8* dst, int width);
-#define HAS_MIRRORROW_UV_NEON
-void MirrorUVRow_NEON(const uint8* src, uint8* dst_a, uint8* dst_b, int width);
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON) || defined(__aarch64__))
 #define HAS_TRANSPOSE_WX8_NEON
 void TransposeWx8_NEON(const uint8* src, int src_stride,
                        uint8* dst, int dst_stride, int width);
@@ -55,7 +51,7 @@ void TransposeUVWx8_NEON(const uint8* src, int src_stride,
                          uint8* dst_a, int dst_stride_a,
                          uint8* dst_b, int dst_stride_b,
                          int width);
-#endif  // defined(__ARM_NEON__)
+#endif
 
 #if !defined(LIBYUV_DISABLE_MIPS) && !defined(__native_client__) && \
     defined(__mips__) && \
@@ -194,31 +190,31 @@ static void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
  convertloop:
     // Read in the data from the source pointer.
     // First round of bit swap.
-    movdqa    xmm0, [eax]
-    movdqa    xmm1, [eax + edi]
+    movdqu    xmm0, [eax]
+    movdqu    xmm1, [eax + edi]
     lea       eax, [eax + 2 * edi]
     movdqa    xmm7, xmm0  // use xmm7 as temp register.
     punpcklbw xmm0, xmm1
     punpckhbw xmm7, xmm1
     movdqa    xmm1, xmm7
-    movdqa    xmm2, [eax]
-    movdqa    xmm3, [eax + edi]
+    movdqu    xmm2, [eax]
+    movdqu    xmm3, [eax + edi]
     lea       eax, [eax + 2 * edi]
     movdqa    xmm7, xmm2
     punpcklbw xmm2, xmm3
     punpckhbw xmm7, xmm3
     movdqa    xmm3, xmm7
-    movdqa    xmm4, [eax]
-    movdqa    xmm5, [eax + edi]
+    movdqu    xmm4, [eax]
+    movdqu    xmm5, [eax + edi]
     lea       eax, [eax + 2 * edi]
     movdqa    xmm7, xmm4
     punpcklbw xmm4, xmm5
     punpckhbw xmm7, xmm5
     movdqa    xmm5, xmm7
-    movdqa    xmm6, [eax]
-    movdqa    xmm7, [eax + edi]
+    movdqu    xmm6, [eax]
+    movdqu    xmm7, [eax + edi]
     lea       eax, [eax + 2 * edi]
-    movdqa    [esp], xmm5  // backup xmm5
+    movdqu    [esp], xmm5  // backup xmm5
     neg       edi
     movdqa    xmm5, xmm6   // use xmm5 as temp register.
     punpcklbw xmm6, xmm7
@@ -239,8 +235,8 @@ static void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
     punpcklwd xmm4, xmm6
     punpckhwd xmm5, xmm6
     movdqa    xmm6, xmm5
-    movdqa    xmm5, [esp]  // restore xmm5
-    movdqa    [esp], xmm6  // backup xmm6
+    movdqu    xmm5, [esp]  // restore xmm5
+    movdqu    [esp], xmm6  // backup xmm6
     movdqa    xmm6, xmm5    // use xmm6 as temp register.
     punpcklwd xmm5, xmm7
     punpckhwd xmm6, xmm7
@@ -251,7 +247,7 @@ static void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
     punpckldq xmm0, xmm4
     punpckhdq xmm6, xmm4
     movdqa    xmm4, xmm6
-    movdqa    xmm6, [esp]  // restore xmm6
+    movdqu    xmm6, [esp]  // restore xmm6
     movlpd    qword ptr [edx], xmm0
     movhpd    qword ptr [ebx], xmm0
     movlpd    qword ptr [edx + esi], xmm4
@@ -296,7 +292,8 @@ static void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
     ret
   }
 }
-#elif !defined(LIBYUV_DISABLE_X86) && \
+#endif
+#if !defined(LIBYUV_DISABLE_X86) && \
     (defined(__i386__) || (defined(__x86_64__) && !defined(__native_client__)))
 #define HAS_TRANSPOSE_WX8_SSSE3
 static void TransposeWx8_SSSE3(const uint8* src, int src_stride,
@@ -379,10 +376,8 @@ static void TransposeWx8_SSSE3(const uint8* src, int src_stride,
       "+r"(width)   // %2
     : "r"((intptr_t)(src_stride)),  // %3
       "r"((intptr_t)(dst_stride))   // %4
-    : "memory", "cc"
-  #if defined(__SSE2__)
-      , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-  #endif
+    : "memory", "cc",
+      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
   );
 }
 
@@ -411,31 +406,31 @@ void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
     "mov    0x2c(%ecx),%ecx                    \n"
 
 "1:                                            \n"
-    "movdqa (%eax),%xmm0                       \n"
-    "movdqa (%eax,%edi,1),%xmm1                \n"
+    "movdqu (%eax),%xmm0                       \n"
+    "movdqu (%eax,%edi,1),%xmm1                \n"
     "lea    (%eax,%edi,2),%eax                 \n"
     "movdqa %xmm0,%xmm7                        \n"
     "punpcklbw %xmm1,%xmm0                     \n"
     "punpckhbw %xmm1,%xmm7                     \n"
     "movdqa %xmm7,%xmm1                        \n"
-    "movdqa (%eax),%xmm2                       \n"
-    "movdqa (%eax,%edi,1),%xmm3                \n"
+    "movdqu (%eax),%xmm2                       \n"
+    "movdqu (%eax,%edi,1),%xmm3                \n"
     "lea    (%eax,%edi,2),%eax                 \n"
     "movdqa %xmm2,%xmm7                        \n"
     "punpcklbw %xmm3,%xmm2                     \n"
     "punpckhbw %xmm3,%xmm7                     \n"
     "movdqa %xmm7,%xmm3                        \n"
-    "movdqa (%eax),%xmm4                       \n"
-    "movdqa (%eax,%edi,1),%xmm5                \n"
+    "movdqu (%eax),%xmm4                       \n"
+    "movdqu (%eax,%edi,1),%xmm5                \n"
     "lea    (%eax,%edi,2),%eax                 \n"
     "movdqa %xmm4,%xmm7                        \n"
     "punpcklbw %xmm5,%xmm4                     \n"
     "punpckhbw %xmm5,%xmm7                     \n"
     "movdqa %xmm7,%xmm5                        \n"
-    "movdqa (%eax),%xmm6                       \n"
-    "movdqa (%eax,%edi,1),%xmm7                \n"
+    "movdqu (%eax),%xmm6                       \n"
+    "movdqu (%eax,%edi,1),%xmm7                \n"
     "lea    (%eax,%edi,2),%eax                 \n"
-    "movdqa %xmm5,(%esp)                       \n"
+    "movdqu %xmm5,(%esp)                       \n"
     "neg    %edi                               \n"
     "movdqa %xmm6,%xmm5                        \n"
     "punpcklbw %xmm7,%xmm6                     \n"
@@ -455,8 +450,8 @@ void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
     "punpcklwd %xmm6,%xmm4                     \n"
     "punpckhwd %xmm6,%xmm5                     \n"
     "movdqa %xmm5,%xmm6                        \n"
-    "movdqa (%esp),%xmm5                       \n"
-    "movdqa %xmm6,(%esp)                       \n"
+    "movdqu (%esp),%xmm5                       \n"
+    "movdqu %xmm6,(%esp)                       \n"
     "movdqa %xmm5,%xmm6                        \n"
     "punpcklwd %xmm7,%xmm5                     \n"
     "punpckhwd %xmm7,%xmm6                     \n"
@@ -465,7 +460,7 @@ void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
     "punpckldq %xmm4,%xmm0                     \n"
     "punpckhdq %xmm4,%xmm6                     \n"
     "movdqa %xmm6,%xmm4                        \n"
-    "movdqa (%esp),%xmm6                       \n"
+    "movdqu (%esp),%xmm6                       \n"
     "movlpd %xmm0,(%edx)                       \n"
     "movhpd %xmm0,(%ebx)                       \n"
     "movlpd %xmm4,(%edx,%esi,1)                \n"
@@ -514,7 +509,8 @@ void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
     "ret                                       \n"
 #endif
 );
-#elif !defined(LIBYUV_DISABLE_X86) && !defined(__native_client__) && \
+#endif
+#if !defined(LIBYUV_DISABLE_X86) && !defined(__native_client__) && \
     defined(__x86_64__)
 // 64 bit version has enough registers to do 16x8 to 8x16 at a time.
 #define HAS_TRANSPOSE_WX8_FAST_SSSE3
@@ -525,38 +521,38 @@ static void TransposeWx8_FAST_SSSE3(const uint8* src, int src_stride,
   // First round of bit swap.
   ".p2align  2                                 \n"
 "1:                                            \n"
-  "movdqa     (%0),%%xmm0                      \n"
-  "movdqa     (%0,%3),%%xmm1                   \n"
+  "movdqu     (%0),%%xmm0                      \n"
+  "movdqu     (%0,%3),%%xmm1                   \n"
   "lea        (%0,%3,2),%0                     \n"
   "movdqa     %%xmm0,%%xmm8                    \n"
   "punpcklbw  %%xmm1,%%xmm0                    \n"
   "punpckhbw  %%xmm1,%%xmm8                    \n"
-  "movdqa     (%0),%%xmm2                      \n"
+  "movdqu     (%0),%%xmm2                      \n"
   "movdqa     %%xmm0,%%xmm1                    \n"
   "movdqa     %%xmm8,%%xmm9                    \n"
   "palignr    $0x8,%%xmm1,%%xmm1               \n"
   "palignr    $0x8,%%xmm9,%%xmm9               \n"
-  "movdqa     (%0,%3),%%xmm3                   \n"
+  "movdqu     (%0,%3),%%xmm3                   \n"
   "lea        (%0,%3,2),%0                     \n"
   "movdqa     %%xmm2,%%xmm10                   \n"
   "punpcklbw  %%xmm3,%%xmm2                    \n"
   "punpckhbw  %%xmm3,%%xmm10                   \n"
   "movdqa     %%xmm2,%%xmm3                    \n"
   "movdqa     %%xmm10,%%xmm11                  \n"
-  "movdqa     (%0),%%xmm4                      \n"
+  "movdqu     (%0),%%xmm4                      \n"
   "palignr    $0x8,%%xmm3,%%xmm3               \n"
   "palignr    $0x8,%%xmm11,%%xmm11             \n"
-  "movdqa     (%0,%3),%%xmm5                   \n"
+  "movdqu     (%0,%3),%%xmm5                   \n"
   "lea        (%0,%3,2),%0                     \n"
   "movdqa     %%xmm4,%%xmm12                   \n"
   "punpcklbw  %%xmm5,%%xmm4                    \n"
   "punpckhbw  %%xmm5,%%xmm12                   \n"
   "movdqa     %%xmm4,%%xmm5                    \n"
   "movdqa     %%xmm12,%%xmm13                  \n"
-  "movdqa     (%0),%%xmm6                      \n"
+  "movdqu     (%0),%%xmm6                      \n"
   "palignr    $0x8,%%xmm5,%%xmm5               \n"
   "palignr    $0x8,%%xmm13,%%xmm13             \n"
-  "movdqa     (%0,%3),%%xmm7                   \n"
+  "movdqu     (%0,%3),%%xmm7                   \n"
   "lea        (%0,%3,2),%0                     \n"
   "movdqa     %%xmm6,%%xmm14                   \n"
   "punpcklbw  %%xmm7,%%xmm6                    \n"
@@ -666,29 +662,29 @@ static void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
   // First round of bit swap.
   ".p2align  2                                 \n"
 "1:                                            \n"
-  "movdqa     (%0),%%xmm0                      \n"
-  "movdqa     (%0,%4),%%xmm1                   \n"
+  "movdqu     (%0),%%xmm0                      \n"
+  "movdqu     (%0,%4),%%xmm1                   \n"
   "lea        (%0,%4,2),%0                     \n"
   "movdqa     %%xmm0,%%xmm8                    \n"
   "punpcklbw  %%xmm1,%%xmm0                    \n"
   "punpckhbw  %%xmm1,%%xmm8                    \n"
   "movdqa     %%xmm8,%%xmm1                    \n"
-  "movdqa     (%0),%%xmm2                      \n"
-  "movdqa     (%0,%4),%%xmm3                   \n"
+  "movdqu     (%0),%%xmm2                      \n"
+  "movdqu     (%0,%4),%%xmm3                   \n"
   "lea        (%0,%4,2),%0                     \n"
   "movdqa     %%xmm2,%%xmm8                    \n"
   "punpcklbw  %%xmm3,%%xmm2                    \n"
   "punpckhbw  %%xmm3,%%xmm8                    \n"
   "movdqa     %%xmm8,%%xmm3                    \n"
-  "movdqa     (%0),%%xmm4                      \n"
-  "movdqa     (%0,%4),%%xmm5                   \n"
+  "movdqu     (%0),%%xmm4                      \n"
+  "movdqu     (%0,%4),%%xmm5                   \n"
   "lea        (%0,%4,2),%0                     \n"
   "movdqa     %%xmm4,%%xmm8                    \n"
   "punpcklbw  %%xmm5,%%xmm4                    \n"
   "punpckhbw  %%xmm5,%%xmm8                    \n"
   "movdqa     %%xmm8,%%xmm5                    \n"
-  "movdqa     (%0),%%xmm6                      \n"
-  "movdqa     (%0,%4),%%xmm7                   \n"
+  "movdqu     (%0),%%xmm6                      \n"
+  "movdqu     (%0,%4),%%xmm7                   \n"
   "lea        (%0,%4,2),%0                     \n"
   "movdqa     %%xmm6,%%xmm8                    \n"
   "punpcklbw  %%xmm7,%%xmm6                    \n"
@@ -818,9 +814,7 @@ void TransposePlane(const uint8* src, int src_stride,
   }
 #endif
 #if defined(HAS_TRANSPOSE_WX8_FAST_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) &&
-      IS_ALIGNED(width, 16) &&
-      IS_ALIGNED(src, 16) && IS_ALIGNED(src_stride, 16)) {
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 16)) {
     TransposeWx8 = TransposeWx8_FAST_SSSE3;
   }
 #endif
@@ -883,29 +877,38 @@ void RotatePlane180(const uint8* src, int src_stride,
   void (*MirrorRow)(const uint8* src, uint8* dst, int width) = MirrorRow_C;
   void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
 #if defined(HAS_MIRRORROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 16)) {
-    MirrorRow = MirrorRow_NEON;
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MirrorRow = MirrorRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorRow = MirrorRow_NEON;
+    }
   }
 #endif
 #if defined(HAS_MIRRORROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16) &&
-      IS_ALIGNED(src, 16) && IS_ALIGNED(src_stride, 16) &&
-      IS_ALIGNED(dst, 16) && IS_ALIGNED(dst_stride, 16)) {
-    MirrorRow = MirrorRow_SSE2;
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    MirrorRow = MirrorRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorRow = MirrorRow_SSE2;
+    }
   }
 #endif
 #if defined(HAS_MIRRORROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 16) &&
-      IS_ALIGNED(src, 16) && IS_ALIGNED(src_stride, 16) &&
-      IS_ALIGNED(dst, 16) && IS_ALIGNED(dst_stride, 16)) {
-    MirrorRow = MirrorRow_SSSE3;
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    MirrorRow = MirrorRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorRow = MirrorRow_SSSE3;
+    }
   }
 #endif
 #if defined(HAS_MIRRORROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 32)) {
-    MirrorRow = MirrorRow_AVX2;
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MirrorRow = MirrorRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      MirrorRow = MirrorRow_AVX2;
+    }
   }
 #endif
+// TODO(fbarchard): Mirror on mips handle unaligned memory.
 #if defined(HAS_MIRRORROW_MIPS_DSPR2)
   if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
       IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4) &&
@@ -913,21 +916,14 @@ void RotatePlane180(const uint8* src, int src_stride,
     MirrorRow = MirrorRow_MIPS_DSPR2;
   }
 #endif
-#if defined(HAS_COPYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 32)) {
-    CopyRow = CopyRow_NEON;
-  }
-#endif
-#if defined(HAS_COPYROW_X86)
-  if (TestCpuFlag(kCpuHasX86) && IS_ALIGNED(width, 4)) {
-    CopyRow = CopyRow_X86;
+#if defined(HAS_COPYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    CopyRow = IS_ALIGNED(width, 32) ? CopyRow_SSE2 : CopyRow_Any_SSE2;
   }
 #endif
-#if defined(HAS_COPYROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 32) &&
-      IS_ALIGNED(src, 16) && IS_ALIGNED(src_stride, 16) &&
-      IS_ALIGNED(dst, 16) && IS_ALIGNED(dst_stride, 16)) {
-    CopyRow = CopyRow_SSE2;
+#if defined(HAS_COPYROW_AVX)
+  if (TestCpuFlag(kCpuHasAVX)) {
+    CopyRow = IS_ALIGNED(width, 64) ? CopyRow_AVX : CopyRow_Any_AVX;
   }
 #endif
 #if defined(HAS_COPYROW_ERMS)
@@ -935,6 +931,11 @@ void RotatePlane180(const uint8* src, int src_stride,
     CopyRow = CopyRow_ERMS;
   }
 #endif
+#if defined(HAS_COPYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    CopyRow = IS_ALIGNED(width, 32) ? CopyRow_NEON : CopyRow_Any_NEON;
+  }
+#endif
 #if defined(HAS_COPYROW_MIPS)
   if (TestCpuFlag(kCpuHasMIPS)) {
     CopyRow = CopyRow_MIPS;
@@ -1010,13 +1011,13 @@ void TransposeUV(const uint8* src, int src_stride,
   if (TestCpuFlag(kCpuHasNEON)) {
     TransposeUVWx8 = TransposeUVWx8_NEON;
   }
-#elif defined(HAS_TRANSPOSE_UVWX8_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) &&
-      IS_ALIGNED(width, 8) &&
-      IS_ALIGNED(src, 16) && IS_ALIGNED(src_stride, 16)) {
+#endif
+#if defined(HAS_TRANSPOSE_UVWX8_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 8)) {
     TransposeUVWx8 = TransposeUVWx8_SSE2;
   }
-#elif defined(HAS_TRANSPOSE_UVWx8_MIPS_DSPR2)
+#endif
+#if defined(HAS_TRANSPOSE_UVWx8_MIPS_DSPR2)
   if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 2) &&
       IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4)) {
     TransposeUVWx8 = TransposeUVWx8_MIPS_DSPR2;
@@ -1084,12 +1085,13 @@ void RotateUV180(const uint8* src, int src_stride,
   if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
     MirrorRowUV = MirrorUVRow_NEON;
   }
-#elif defined(HAS_MIRRORROW_UV_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 16) &&
-      IS_ALIGNED(src, 16) && IS_ALIGNED(src_stride, 16)) {
+#endif
+#if defined(HAS_MIRRORROW_UV_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 16)) {
     MirrorRowUV = MirrorUVRow_SSSE3;
   }
-#elif defined(HAS_MIRRORUVROW_MIPS_DSPR2)
+#endif
+#if defined(HAS_MIRRORUVROW_MIPS_DSPR2)
   if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
       IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4)) {
     MirrorRowUV = MirrorUVRow_MIPS_DSPR2;
diff --git a/third_party/libyuv/source/rotate_argb.cc b/third_party/libyuv/source/rotate_argb.cc
index ab0f9ce07..b9673db15 100644
--- a/third_party/libyuv/source/rotate_argb.cc
+++ b/third_party/libyuv/source/rotate_argb.cc
@@ -31,7 +31,7 @@ void ScaleARGBRowDownEven_SSE2(const uint8* src_ptr, int src_stride,
                                uint8* dst_ptr, int dst_width);
 #endif
 #if !defined(LIBYUV_DISABLE_NEON) && !defined(__native_client__) && \
-    (defined(__ARM_NEON__) || defined(LIBYUV_NEON))
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON) || defined(__aarch64__))
 #define HAS_SCALEARGBROWDOWNEVEN_NEON
 void ScaleARGBRowDownEven_NEON(const uint8* src_ptr, int src_stride,
                                int src_stepx,
@@ -50,13 +50,12 @@ static void ARGBTranspose(const uint8* src, int src_stride,
   void (*ScaleARGBRowDownEven)(const uint8* src_ptr, int src_stride,
       int src_step, uint8* dst_ptr, int dst_width) = ScaleARGBRowDownEven_C;
 #if defined(HAS_SCALEARGBROWDOWNEVEN_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(height, 4) &&  // Width of dest.
-      IS_ALIGNED(dst, 16) && IS_ALIGNED(dst_stride, 16)) {
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(height, 4)) {  // Width of dest.
     ScaleARGBRowDownEven = ScaleARGBRowDownEven_SSE2;
   }
-#elif defined(HAS_SCALEARGBROWDOWNEVEN_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(height, 4) &&  // Width of dest.
-      IS_ALIGNED(src, 4)) {
+#endif
+#if defined(HAS_SCALEARGBROWDOWNEVEN_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(height, 4)) {  // Width of dest.
     ScaleARGBRowDownEven = ScaleARGBRowDownEven_NEON;
   }
 #endif
@@ -102,38 +101,38 @@ void ARGBRotate180(const uint8* src, int src_stride,
   void (*ARGBMirrorRow)(const uint8* src, uint8* dst, int width) =
       ARGBMirrorRow_C;
   void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
-#if defined(HAS_ARGBMIRRORROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 4) &&
-      IS_ALIGNED(src, 16) && IS_ALIGNED(src_stride, 16) &&
-      IS_ALIGNED(dst, 16) && IS_ALIGNED(dst_stride, 16)) {
-    ARGBMirrorRow = ARGBMirrorRow_SSSE3;
-  }
-#endif
-#if defined(HAS_ARGBMIRRORROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 8)) {
-    ARGBMirrorRow = ARGBMirrorRow_AVX2;
-  }
-#endif
 #if defined(HAS_ARGBMIRRORROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 4)) {
-    ARGBMirrorRow = ARGBMirrorRow_NEON;
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_NEON;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBMirrorRow = ARGBMirrorRow_NEON;
+    }
   }
 #endif
-#if defined(HAS_COPYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width * 4, 32)) {
-    CopyRow = CopyRow_NEON;
+#if defined(HAS_ARGBMIRRORROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBMirrorRow = ARGBMirrorRow_SSE2;
+    }
   }
 #endif
-#if defined(HAS_COPYROW_X86)
-  if (TestCpuFlag(kCpuHasX86)) {
-    CopyRow = CopyRow_X86;
+#if defined(HAS_ARGBMIRRORROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBMirrorRow = ARGBMirrorRow_AVX2;
+    }
   }
 #endif
 #if defined(HAS_COPYROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width * 4, 32) &&
-      IS_ALIGNED(src, 16) && IS_ALIGNED(src_stride, 16) &&
-      IS_ALIGNED(dst, 16) && IS_ALIGNED(dst_stride, 16)) {
-    CopyRow = CopyRow_SSE2;
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    CopyRow = IS_ALIGNED(width * 4, 32) ? CopyRow_SSE2 : CopyRow_Any_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_AVX)
+  if (TestCpuFlag(kCpuHasAVX)) {
+    CopyRow = IS_ALIGNED(width * 4, 64) ? CopyRow_AVX : CopyRow_Any_AVX;
   }
 #endif
 #if defined(HAS_COPYROW_ERMS)
@@ -141,6 +140,11 @@ void ARGBRotate180(const uint8* src, int src_stride,
     CopyRow = CopyRow_ERMS;
   }
 #endif
+#if defined(HAS_COPYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    CopyRow = IS_ALIGNED(width * 4, 32) ? CopyRow_NEON : CopyRow_Any_NEON;
+  }
+#endif
 #if defined(HAS_COPYROW_MIPS)
   if (TestCpuFlag(kCpuHasMIPS)) {
     CopyRow = CopyRow_MIPS;
diff --git a/third_party/libyuv/source/rotate_neon.cc b/third_party/libyuv/source/rotate_neon.cc
index d354e11fa..a23a40fee 100644
--- a/third_party/libyuv/source/rotate_neon.cc
+++ b/third_party/libyuv/source/rotate_neon.cc
@@ -17,7 +17,8 @@ namespace libyuv {
 extern "C" {
 #endif
 
-#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__)
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \
+    !defined(__aarch64__)
 
 static uvec8 kVTbl4x4Transpose =
   { 0,  4,  8, 12,  1,  5,  9, 13,  2,  6, 10, 14,  3,  7, 11, 15 };
@@ -525,7 +526,7 @@ void TransposeUVWx8_NEON(const uint8* src, int src_stride,
       "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11"
   );
 }
-#endif
+#endif  // defined(__ARM_NEON__) && !defined(__aarch64__)
 
 #ifdef __cplusplus
 }  // extern "C"
diff --git a/third_party/libyuv/source/rotate_neon64.cc b/third_party/libyuv/source/rotate_neon64.cc
new file mode 100644
index 000000000..92358af7f
--- /dev/null
+++ b/third_party/libyuv/source/rotate_neon64.cc
@@ -0,0 +1,543 @@
+/*
+ *  Copyright 2014 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC Neon armv8 64 bit.
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+static uvec8 kVTbl4x4Transpose =
+  { 0,  4,  8, 12,  1,  5,  9, 13,  2,  6, 10, 14,  3,  7, 11, 15 };
+
+void TransposeWx8_NEON(const uint8* src, int src_stride,
+                       uint8* dst, int dst_stride,
+                       int width) {
+  const uint8* src_temp = NULL;
+  int64 width64 = (int64) width;  // Work around clang 3.4 warning.
+  asm volatile (
+    // loops are on blocks of 8. loop will stop when
+    // counter gets to or below 0. starting the counter
+    // at w-8 allow for this
+    "sub         %3, %3, #8                      \n"
+
+    // handle 8x8 blocks. this should be the majority of the plane
+    "1:                                          \n"
+      "mov         %0, %1                        \n"
+
+      MEMACCESS(0)
+      "ld1        {v0.8b}, [%0], %5              \n"
+      MEMACCESS(0)
+      "ld1        {v1.8b}, [%0], %5              \n"
+      MEMACCESS(0)
+      "ld1        {v2.8b}, [%0], %5              \n"
+      MEMACCESS(0)
+      "ld1        {v3.8b}, [%0], %5              \n"
+      MEMACCESS(0)
+      "ld1        {v4.8b}, [%0], %5              \n"
+      MEMACCESS(0)
+      "ld1        {v5.8b}, [%0], %5              \n"
+      MEMACCESS(0)
+      "ld1        {v6.8b}, [%0], %5              \n"
+      MEMACCESS(0)
+      "ld1        {v7.8b}, [%0]                  \n"
+
+      "trn2     v16.8b, v0.8b, v1.8b             \n"
+      "trn1     v17.8b, v0.8b, v1.8b             \n"
+      "trn2     v18.8b, v2.8b, v3.8b             \n"
+      "trn1     v19.8b, v2.8b, v3.8b             \n"
+      "trn2     v20.8b, v4.8b, v5.8b             \n"
+      "trn1     v21.8b, v4.8b, v5.8b             \n"
+      "trn2     v22.8b, v6.8b, v7.8b             \n"
+      "trn1     v23.8b, v6.8b, v7.8b             \n"
+
+      "trn2     v3.4h, v17.4h, v19.4h            \n"
+      "trn1     v1.4h, v17.4h, v19.4h            \n"
+      "trn2     v2.4h, v16.4h, v18.4h            \n"
+      "trn1     v0.4h, v16.4h, v18.4h            \n"
+      "trn2     v7.4h, v21.4h, v23.4h            \n"
+      "trn1     v5.4h, v21.4h, v23.4h            \n"
+      "trn2     v6.4h, v20.4h, v22.4h            \n"
+      "trn1     v4.4h, v20.4h, v22.4h            \n"
+
+      "trn2     v21.2s, v1.2s, v5.2s             \n"
+      "trn1     v17.2s, v1.2s, v5.2s             \n"
+      "trn2     v20.2s, v0.2s, v4.2s             \n"
+      "trn1     v16.2s, v0.2s, v4.2s             \n"
+      "trn2     v23.2s, v3.2s, v7.2s             \n"
+      "trn1     v19.2s, v3.2s, v7.2s             \n"
+      "trn2     v22.2s, v2.2s, v6.2s             \n"
+      "trn1     v18.2s, v2.2s, v6.2s             \n"
+
+      "mov         %0, %2                        \n"
+
+    MEMACCESS(0)
+      "st1      {v17.8b}, [%0], %6               \n"
+    MEMACCESS(0)
+      "st1      {v16.8b}, [%0], %6               \n"
+    MEMACCESS(0)
+      "st1      {v19.8b}, [%0], %6               \n"
+    MEMACCESS(0)
+      "st1      {v18.8b}, [%0], %6               \n"
+    MEMACCESS(0)
+      "st1      {v21.8b}, [%0], %6               \n"
+    MEMACCESS(0)
+      "st1      {v20.8b}, [%0], %6               \n"
+    MEMACCESS(0)
+      "st1      {v23.8b}, [%0], %6               \n"
+    MEMACCESS(0)
+      "st1      {v22.8b}, [%0]                   \n"
+
+      "add         %1, %1, #8                    \n"  // src += 8
+      "add         %2, %2, %6, lsl #3            \n"  // dst += 8 * dst_stride
+      "subs        %3, %3, #8                    \n"  // w   -= 8
+      "b.ge        1b                            \n"
+
+    // add 8 back to counter. if the result is 0 there are
+    // no residuals.
+    "adds        %3, %3, #8                      \n"
+    "b.eq        4f                              \n"
+
+    // some residual, so between 1 and 7 lines left to transpose
+    "cmp         %3, #2                          \n"
+    "b.lt        3f                              \n"
+
+    "cmp         %3, #4                          \n"
+    "b.lt        2f                              \n"
+
+    // 4x8 block
+    "mov         %0, %1                          \n"
+    MEMACCESS(0)
+    "ld1     {v0.s}[0], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v0.s}[1], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v0.s}[2], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v0.s}[3], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v1.s}[0], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v1.s}[1], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v1.s}[2], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v1.s}[3], [%0]                     \n"
+
+    "mov         %0, %2                          \n"
+
+    MEMACCESS(4)
+    "ld1      {v2.16b}, [%4]                     \n"
+
+    "tbl      v3.16b, {v0.16b}, v2.16b           \n"
+    "tbl      v0.16b, {v1.16b}, v2.16b           \n"
+
+    // TODO(frkoenig): Rework shuffle above to
+    // write out with 4 instead of 8 writes.
+    MEMACCESS(0)
+    "st1 {v3.s}[0], [%0], %6                     \n"
+    MEMACCESS(0)
+    "st1 {v3.s}[1], [%0], %6                     \n"
+    MEMACCESS(0)
+    "st1 {v3.s}[2], [%0], %6                     \n"
+    MEMACCESS(0)
+    "st1 {v3.s}[3], [%0]                         \n"
+
+    "add         %0, %2, #4                      \n"
+    MEMACCESS(0)
+    "st1 {v0.s}[0], [%0], %6                     \n"
+    MEMACCESS(0)
+    "st1 {v0.s}[1], [%0], %6                     \n"
+    MEMACCESS(0)
+    "st1 {v0.s}[2], [%0], %6                     \n"
+    MEMACCESS(0)
+    "st1 {v0.s}[3], [%0]                         \n"
+
+    "add         %1, %1, #4                      \n"  // src += 4
+    "add         %2, %2, %6, lsl #2              \n"  // dst += 4 * dst_stride
+    "subs        %3, %3, #4                      \n"  // w   -= 4
+    "b.eq        4f                              \n"
+
+    // some residual, check to see if it includes a 2x8 block,
+    // or less
+    "cmp         %3, #2                          \n"
+    "b.lt        3f                              \n"
+
+    // 2x8 block
+    "2:                                          \n"
+    "mov         %0, %1                          \n"
+    MEMACCESS(0)
+    "ld1     {v0.h}[0], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v1.h}[0], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v0.h}[1], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v1.h}[1], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v0.h}[2], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v1.h}[2], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v0.h}[3], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v1.h}[3], [%0]                     \n"
+
+    "trn2    v2.8b, v0.8b, v1.8b                 \n"
+    "trn1    v3.8b, v0.8b, v1.8b                 \n"
+
+    "mov         %0, %2                          \n"
+
+    MEMACCESS(0)
+    "st1     {v3.8b}, [%0], %6                   \n"
+    MEMACCESS(0)
+    "st1     {v2.8b}, [%0]                       \n"
+
+    "add         %1, %1, #2                      \n"  // src += 2
+    "add         %2, %2, %6, lsl #1              \n"  // dst += 2 * dst_stride
+    "subs        %3, %3,  #2                     \n"  // w   -= 2
+    "b.eq        4f                              \n"
+
+    // 1x8 block
+    "3:                                          \n"
+    MEMACCESS(1)
+    "ld1         {v0.b}[0], [%1], %5             \n"
+    MEMACCESS(1)
+    "ld1         {v0.b}[1], [%1], %5             \n"
+    MEMACCESS(1)
+    "ld1         {v0.b}[2], [%1], %5             \n"
+    MEMACCESS(1)
+    "ld1         {v0.b}[3], [%1], %5             \n"
+    MEMACCESS(1)
+    "ld1         {v0.b}[4], [%1], %5             \n"
+    MEMACCESS(1)
+    "ld1         {v0.b}[5], [%1], %5             \n"
+    MEMACCESS(1)
+    "ld1         {v0.b}[6], [%1], %5             \n"
+    MEMACCESS(1)
+    "ld1         {v0.b}[7], [%1]                 \n"
+
+    MEMACCESS(2)
+    "st1         {v0.8b}, [%2]                   \n"
+
+    "4:                                          \n"
+
+    : "+r"(src_temp),                             // %0
+      "+r"(src),                                  // %1
+      "+r"(dst),                                  // %2
+      "+r"(width64)                               // %3
+    : "r"(&kVTbl4x4Transpose),                    // %4
+      "r"(static_cast<ptrdiff_t>(src_stride)),    // %5
+      "r"(static_cast<ptrdiff_t>(dst_stride))     // %6
+    : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16",
+      "v17", "v18", "v19", "v20", "v21", "v22", "v23"
+  );
+}
+
+static uint8 kVTbl4x4TransposeDi[32] =
+  { 0,  16, 32, 48,  2, 18, 34, 50,  4, 20, 36, 52,  6, 22, 38, 54,
+    1,  17, 33, 49,  3, 19, 35, 51,  5, 21, 37, 53,  7, 23, 39, 55};
+
+void TransposeUVWx8_NEON(const uint8* src, int src_stride,
+                         uint8* dst_a, int dst_stride_a,
+                         uint8* dst_b, int dst_stride_b,
+                         int width) {
+  const uint8* src_temp = NULL;
+  int64 width64 = (int64) width;  // Work around clang 3.4 warning.
+  asm volatile (
+    // loops are on blocks of 8. loop will stop when
+    // counter gets to or below 0. starting the counter
+    // at w-8 allow for this
+    "sub       %4, %4, #8                      \n"
+
+    // handle 8x8 blocks. this should be the majority of the plane
+    "1:                                        \n"
+    "mov       %0, %1                          \n"
+
+    MEMACCESS(0)
+    "ld1       {v0.16b}, [%0], %5              \n"
+    MEMACCESS(0)
+    "ld1       {v1.16b}, [%0], %5              \n"
+    MEMACCESS(0)
+    "ld1       {v2.16b}, [%0], %5              \n"
+    MEMACCESS(0)
+    "ld1       {v3.16b}, [%0], %5              \n"
+    MEMACCESS(0)
+    "ld1       {v4.16b}, [%0], %5              \n"
+    MEMACCESS(0)
+    "ld1       {v5.16b}, [%0], %5              \n"
+    MEMACCESS(0)
+    "ld1       {v6.16b}, [%0], %5              \n"
+    MEMACCESS(0)
+    "ld1       {v7.16b}, [%0]                  \n"
+
+    "trn1      v16.16b, v0.16b, v1.16b         \n"
+    "trn2      v17.16b, v0.16b, v1.16b         \n"
+    "trn1      v18.16b, v2.16b, v3.16b         \n"
+    "trn2      v19.16b, v2.16b, v3.16b         \n"
+    "trn1      v20.16b, v4.16b, v5.16b         \n"
+    "trn2      v21.16b, v4.16b, v5.16b         \n"
+    "trn1      v22.16b, v6.16b, v7.16b         \n"
+    "trn2      v23.16b, v6.16b, v7.16b         \n"
+
+    "trn1      v0.8h, v16.8h, v18.8h           \n"
+    "trn2      v1.8h, v16.8h, v18.8h           \n"
+    "trn1      v2.8h, v20.8h, v22.8h           \n"
+    "trn2      v3.8h, v20.8h, v22.8h           \n"
+    "trn1      v4.8h, v17.8h, v19.8h           \n"
+    "trn2      v5.8h, v17.8h, v19.8h           \n"
+    "trn1      v6.8h, v21.8h, v23.8h           \n"
+    "trn2      v7.8h, v21.8h, v23.8h           \n"
+
+    "trn1      v16.4s, v0.4s, v2.4s            \n"
+    "trn2      v17.4s, v0.4s, v2.4s            \n"
+    "trn1      v18.4s, v1.4s, v3.4s            \n"
+    "trn2      v19.4s, v1.4s, v3.4s            \n"
+    "trn1      v20.4s, v4.4s, v6.4s            \n"
+    "trn2      v21.4s, v4.4s, v6.4s            \n"
+    "trn1      v22.4s, v5.4s, v7.4s            \n"
+    "trn2      v23.4s, v5.4s, v7.4s            \n"
+
+    "mov       %0, %2                          \n"
+
+    MEMACCESS(0)
+    "st1       {v16.d}[0], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v18.d}[0], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v17.d}[0], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v19.d}[0], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v16.d}[1], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v18.d}[1], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v17.d}[1], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v19.d}[1], [%0]                \n"
+
+    "mov       %0, %3                          \n"
+
+    MEMACCESS(0)
+    "st1       {v20.d}[0], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v22.d}[0], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v21.d}[0], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v23.d}[0], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v20.d}[1], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v22.d}[1], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v21.d}[1], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v23.d}[1], [%0]                \n"
+
+    "add       %1, %1, #16                     \n"  // src   += 8*2
+    "add       %2, %2, %6, lsl #3              \n"  // dst_a += 8 * dst_stride_a
+    "add       %3, %3, %7, lsl #3              \n"  // dst_b += 8 * dst_stride_b
+    "subs      %4, %4,  #8                     \n"  // w     -= 8
+    "b.ge      1b                              \n"
+
+    // add 8 back to counter. if the result is 0 there are
+    // no residuals.
+    "adds      %4, %4, #8                      \n"
+    "b.eq      4f                              \n"
+
+    // some residual, so between 1 and 7 lines left to transpose
+    "cmp       %4, #2                          \n"
+    "b.lt      3f                              \n"
+
+    "cmp       %4, #4                          \n"
+    "b.lt      2f                              \n"
+
+    // TODO(frkoenig): Clean this up
+    // 4x8 block
+    "mov       %0, %1                          \n"
+    MEMACCESS(0)
+    "ld1       {v0.8b}, [%0], %5               \n"
+    MEMACCESS(0)
+    "ld1       {v1.8b}, [%0], %5               \n"
+    MEMACCESS(0)
+    "ld1       {v2.8b}, [%0], %5               \n"
+    MEMACCESS(0)
+    "ld1       {v3.8b}, [%0], %5               \n"
+    MEMACCESS(0)
+    "ld1       {v4.8b}, [%0], %5               \n"
+    MEMACCESS(0)
+    "ld1       {v5.8b}, [%0], %5               \n"
+    MEMACCESS(0)
+    "ld1       {v6.8b}, [%0], %5               \n"
+    MEMACCESS(0)
+    "ld1       {v7.8b}, [%0]                   \n"
+
+    MEMACCESS(8)
+    "ld1       {v30.16b}, [%8], #16            \n"
+    "ld1       {v31.16b}, [%8]                 \n"
+
+    "tbl       v16.16b, {v0.16b, v1.16b, v2.16b, v3.16b}, v30.16b  \n"
+    "tbl       v17.16b, {v0.16b, v1.16b, v2.16b, v3.16b}, v31.16b  \n"
+    "tbl       v18.16b, {v4.16b, v5.16b, v6.16b, v7.16b}, v30.16b  \n"
+    "tbl       v19.16b, {v4.16b, v5.16b, v6.16b, v7.16b}, v31.16b  \n"
+
+    "mov       %0, %2                          \n"
+
+    MEMACCESS(0)
+    "st1       {v16.s}[0],  [%0], %6           \n"
+    MEMACCESS(0)
+    "st1       {v16.s}[1],  [%0], %6           \n"
+    MEMACCESS(0)
+    "st1       {v16.s}[2],  [%0], %6           \n"
+    MEMACCESS(0)
+    "st1       {v16.s}[3],  [%0], %6           \n"
+
+    "add       %0, %2, #4                      \n"
+    MEMACCESS(0)
+    "st1       {v18.s}[0], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v18.s}[1], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v18.s}[2], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v18.s}[3], [%0]                \n"
+
+    "mov       %0, %3                          \n"
+
+    MEMACCESS(0)
+    "st1       {v17.s}[0], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v17.s}[1], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v17.s}[2], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v17.s}[3], [%0], %7            \n"
+
+    "add       %0, %3, #4                      \n"
+    MEMACCESS(0)
+    "st1       {v19.s}[0],  [%0], %7           \n"
+    MEMACCESS(0)
+    "st1       {v19.s}[1],  [%0], %7           \n"
+    MEMACCESS(0)
+    "st1       {v19.s}[2],  [%0], %7           \n"
+    MEMACCESS(0)
+    "st1       {v19.s}[3],  [%0]               \n"
+
+    "add       %1, %1, #8                      \n"  // src   += 4 * 2
+    "add       %2, %2, %6, lsl #2              \n"  // dst_a += 4 * dst_stride_a
+    "add       %3, %3, %7, lsl #2              \n"  // dst_b += 4 * dst_stride_b
+    "subs      %4,  %4,  #4                    \n"  // w     -= 4
+    "b.eq      4f                              \n"
+
+    // some residual, check to see if it includes a 2x8 block,
+    // or less
+    "cmp       %4, #2                          \n"
+    "b.lt      3f                              \n"
+
+    // 2x8 block
+    "2:                                        \n"
+    "mov       %0, %1                          \n"
+    MEMACCESS(0)
+    "ld2       {v0.h, v1.h}[0], [%0], %5       \n"
+    MEMACCESS(0)
+    "ld2       {v2.h, v3.h}[0], [%0], %5       \n"
+    MEMACCESS(0)
+    "ld2       {v0.h, v1.h}[1], [%0], %5       \n"
+    MEMACCESS(0)
+    "ld2       {v2.h, v3.h}[1], [%0], %5       \n"
+    MEMACCESS(0)
+    "ld2       {v0.h, v1.h}[2], [%0], %5       \n"
+    MEMACCESS(0)
+    "ld2       {v2.h, v3.h}[2], [%0], %5       \n"
+    MEMACCESS(0)
+    "ld2       {v0.h, v1.h}[3], [%0], %5       \n"
+    MEMACCESS(0)
+    "ld2       {v2.h, v3.h}[3], [%0]           \n"
+
+    "trn1      v4.8b, v0.8b, v2.8b             \n"
+    "trn2      v5.8b, v0.8b, v2.8b             \n"
+    "trn1      v6.8b, v1.8b, v3.8b             \n"
+    "trn2      v7.8b, v1.8b, v3.8b             \n"
+
+    "mov       %0, %2                          \n"
+
+    MEMACCESS(0)
+    "st1       {v4.d}[0], [%0], %6             \n"
+    MEMACCESS(0)
+    "st1       {v6.d}[0], [%0]                 \n"
+
+    "mov       %0, %3                          \n"
+
+    MEMACCESS(0)
+    "st1       {v5.d}[0], [%0], %7             \n"
+    MEMACCESS(0)
+    "st1       {v7.d}[0], [%0]                 \n"
+
+    "add       %1, %1, #4                      \n"  // src   += 2 * 2
+    "add       %2, %2, %6, lsl #1              \n"  // dst_a += 2 * dst_stride_a
+    "add       %3, %3, %7, lsl #1              \n"  // dst_b += 2 * dst_stride_b
+    "subs      %4,  %4,  #2                    \n"  // w     -= 2
+    "b.eq      4f                              \n"
+
+    // 1x8 block
+    "3:                                        \n"
+    MEMACCESS(1)
+    "ld2       {v0.b, v1.b}[0], [%1], %5       \n"
+    MEMACCESS(1)
+    "ld2       {v0.b, v1.b}[1], [%1], %5       \n"
+    MEMACCESS(1)
+    "ld2       {v0.b, v1.b}[2], [%1], %5       \n"
+    MEMACCESS(1)
+    "ld2       {v0.b, v1.b}[3], [%1], %5       \n"
+    MEMACCESS(1)
+    "ld2       {v0.b, v1.b}[4], [%1], %5       \n"
+    MEMACCESS(1)
+    "ld2       {v0.b, v1.b}[5], [%1], %5       \n"
+    MEMACCESS(1)
+    "ld2       {v0.b, v1.b}[6], [%1], %5       \n"
+    MEMACCESS(1)
+    "ld2       {v0.b, v1.b}[7], [%1]           \n"
+
+    MEMACCESS(2)
+    "st1       {v0.d}[0], [%2]                 \n"
+    MEMACCESS(3)
+    "st1       {v1.d}[0], [%3]                 \n"
+
+    "4:                                        \n"
+
+    : "+r"(src_temp),                             // %0
+      "+r"(src),                                  // %1
+      "+r"(dst_a),                                // %2
+      "+r"(dst_b),                                // %3
+      "+r"(width64)                               // %4
+    : "r"(static_cast<ptrdiff_t>(src_stride)),    // %5
+      "r"(static_cast<ptrdiff_t>(dst_stride_a)),  // %6
+      "r"(static_cast<ptrdiff_t>(dst_stride_b)),  // %7
+      "r"(&kVTbl4x4TransposeDi)                   // %8
+    : "memory", "cc",
+      "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+      "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23",
+      "v30", "v31"
+  );
+}
+#endif  // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/libyuv/source/row_any.cc b/third_party/libyuv/source/row_any.cc
index ce8b3dad1..19340b3b7 100644
--- a/third_party/libyuv/source/row_any.cc
+++ b/third_party/libyuv/source/row_any.cc
@@ -17,17 +17,14 @@ namespace libyuv {
 extern "C" {
 #endif
 
-// TODO(fbarchard): Consider 'any' functions handling any quantity of pixels.
-// TODO(fbarchard): Consider 'any' functions handling odd alignment.
 // YUV to RGB does multiple of 8 with SIMD and remainder with C.
 #define YANY(NAMEANY, I420TORGB_SIMD, I420TORGB_C, UV_SHIFT, BPP, MASK)        \
-    void NAMEANY(const uint8* y_buf,                                           \
-                 const uint8* u_buf,                                           \
-                 const uint8* v_buf,                                           \
-                 uint8* rgb_buf,                                               \
-                 int width) {                                                  \
+    void NAMEANY(const uint8* y_buf, const uint8* u_buf, const uint8* v_buf,   \
+                 uint8* rgb_buf, int width) {                                  \
       int n = width & ~MASK;                                                   \
-      I420TORGB_SIMD(y_buf, u_buf, v_buf, rgb_buf, n);                         \
+      if (n > 0) {                                                             \
+        I420TORGB_SIMD(y_buf, u_buf, v_buf, rgb_buf, n);                       \
+      }                                                                        \
       I420TORGB_C(y_buf + n,                                                   \
                   u_buf + (n >> UV_SHIFT),                                     \
                   v_buf + (n >> UV_SHIFT),                                     \
@@ -35,36 +32,59 @@ extern "C" {
     }
 
 #ifdef HAS_I422TOARGBROW_SSSE3
-YANY(I422ToARGBRow_Any_SSSE3, I422ToARGBRow_Unaligned_SSSE3, I422ToARGBRow_C,
+YANY(I422ToARGBRow_Any_SSSE3, I422ToARGBRow_SSSE3, I422ToARGBRow_C,
      1, 4, 7)
-#endif  // HAS_I422TOARGBROW_SSSE3
+#endif
 #ifdef HAS_I444TOARGBROW_SSSE3
-YANY(I444ToARGBRow_Any_SSSE3, I444ToARGBRow_Unaligned_SSSE3, I444ToARGBRow_C,
+YANY(I444ToARGBRow_Any_SSSE3, I444ToARGBRow_SSSE3, I444ToARGBRow_C,
      0, 4, 7)
-YANY(I411ToARGBRow_Any_SSSE3, I411ToARGBRow_Unaligned_SSSE3, I411ToARGBRow_C,
+YANY(I411ToARGBRow_Any_SSSE3, I411ToARGBRow_SSSE3, I411ToARGBRow_C,
      2, 4, 7)
-YANY(I422ToBGRARow_Any_SSSE3, I422ToBGRARow_Unaligned_SSSE3, I422ToBGRARow_C,
+YANY(I422ToBGRARow_Any_SSSE3, I422ToBGRARow_SSSE3, I422ToBGRARow_C,
      1, 4, 7)
-YANY(I422ToABGRRow_Any_SSSE3, I422ToABGRRow_Unaligned_SSSE3, I422ToABGRRow_C,
+YANY(I422ToABGRRow_Any_SSSE3, I422ToABGRRow_SSSE3, I422ToABGRRow_C,
      1, 4, 7)
-YANY(I422ToRGBARow_Any_SSSE3, I422ToRGBARow_Unaligned_SSSE3, I422ToRGBARow_C,
+YANY(I422ToRGBARow_Any_SSSE3, I422ToRGBARow_SSSE3, I422ToRGBARow_C,
      1, 4, 7)
-// I422ToRGB565Row_SSSE3 is unaligned.
 YANY(I422ToARGB4444Row_Any_SSSE3, I422ToARGB4444Row_SSSE3, I422ToARGB4444Row_C,
      1, 2, 7)
 YANY(I422ToARGB1555Row_Any_SSSE3, I422ToARGB1555Row_SSSE3, I422ToARGB1555Row_C,
      1, 2, 7)
 YANY(I422ToRGB565Row_Any_SSSE3, I422ToRGB565Row_SSSE3, I422ToRGB565Row_C,
      1, 2, 7)
-// I422ToRGB24Row_SSSE3 is unaligned.
 YANY(I422ToRGB24Row_Any_SSSE3, I422ToRGB24Row_SSSE3, I422ToRGB24Row_C, 1, 3, 7)
 YANY(I422ToRAWRow_Any_SSSE3, I422ToRAWRow_SSSE3, I422ToRAWRow_C, 1, 3, 7)
 YANY(I422ToYUY2Row_Any_SSE2, I422ToYUY2Row_SSE2, I422ToYUY2Row_C, 1, 2, 15)
 YANY(I422ToUYVYRow_Any_SSE2, I422ToUYVYRow_SSE2, I422ToUYVYRow_C, 1, 2, 15)
 #endif  // HAS_I444TOARGBROW_SSSE3
+#ifdef HAS_J422TOARGBROW_SSSE3
+YANY(J422ToARGBRow_Any_SSSE3, J422ToARGBRow_SSSE3, J422ToARGBRow_C,
+     1, 4, 7)
+#endif
 #ifdef HAS_I422TOARGBROW_AVX2
 YANY(I422ToARGBRow_Any_AVX2, I422ToARGBRow_AVX2, I422ToARGBRow_C, 1, 4, 15)
-#endif  // HAS_I422TOARGBROW_AVX2
+#endif
+#ifdef HAS_I422TOBGRAROW_AVX2
+YANY(I422ToBGRARow_Any_AVX2, I422ToBGRARow_AVX2, I422ToBGRARow_C, 1, 4, 15)
+#endif
+#ifdef HAS_I422TORGBAROW_AVX2
+YANY(I422ToRGBARow_Any_AVX2, I422ToRGBARow_AVX2, I422ToRGBARow_C, 1, 4, 15)
+#endif
+#ifdef HAS_I422TOABGRROW_AVX2
+YANY(I422ToABGRRow_Any_AVX2, I422ToABGRRow_AVX2, I422ToABGRRow_C, 1, 4, 15)
+#endif
+#ifdef HAS_I422TOARGB4444ROW_AVX2
+YANY(I422ToARGB4444Row_Any_AVX2, I422ToARGB4444Row_AVX2, I422ToARGB4444Row_C,
+     1, 2, 7)
+#endif
+#ifdef HAS_I422TOARGB1555ROW_AVX2
+YANY(I422ToARGB1555Row_Any_AVX2, I422ToARGB1555Row_AVX2, I422ToARGB1555Row_C,
+     1, 2, 7)
+#endif
+#ifdef HAS_I422TORGB565ROW_AVX2
+YANY(I422ToRGB565Row_Any_AVX2, I422ToRGB565Row_AVX2, I422ToRGB565Row_C,
+     1, 2, 7)
+#endif
 #ifdef HAS_I422TOARGBROW_NEON
 YANY(I444ToARGBRow_Any_NEON, I444ToARGBRow_NEON, I444ToARGBRow_C, 0, 4, 7)
 YANY(I422ToARGBRow_Any_NEON, I422ToARGBRow_NEON, I422ToARGBRow_C, 1, 4, 7)
@@ -79,214 +99,240 @@ YANY(I422ToARGB4444Row_Any_NEON, I422ToARGB4444Row_NEON, I422ToARGB4444Row_C,
 YANY(I422ToARGB1555Row_Any_NEON, I422ToARGB1555Row_NEON, I422ToARGB1555Row_C,
      1, 2, 7)
 YANY(I422ToRGB565Row_Any_NEON, I422ToRGB565Row_NEON, I422ToRGB565Row_C, 1, 2, 7)
-#endif  // HAS_I422TOARGBROW_NEON
+#endif
 #ifdef HAS_I422TOYUY2ROW_NEON
 YANY(I422ToYUY2Row_Any_NEON, I422ToYUY2Row_NEON, I422ToYUY2Row_C, 1, 2, 15)
-#endif  // HAS_I422TOYUY2ROW_NEON
+#endif
 #ifdef HAS_I422TOUYVYROW_NEON
 YANY(I422ToUYVYRow_Any_NEON, I422ToUYVYRow_NEON, I422ToUYVYRow_C, 1, 2, 15)
-#endif  // HAS_I422TOUYVYROW_NEON
+#endif
 #undef YANY
 
 // Wrappers to handle odd width
-#define NV2NY(NAMEANY, NV12TORGB_SIMD, NV12TORGB_C, UV_SHIFT, BPP)             \
-    void NAMEANY(const uint8* y_buf,                                           \
-                 const uint8* uv_buf,                                          \
-                 uint8* rgb_buf,                                               \
-                 int width) {                                                  \
-      int n = width & ~7;                                                      \
-      NV12TORGB_SIMD(y_buf, uv_buf, rgb_buf, n);                               \
+#define NV2NY(NAMEANY, NV12TORGB_SIMD, NV12TORGB_C, UV_SHIFT, BPP, MASK)       \
+    void NAMEANY(const uint8* y_buf, const uint8* uv_buf,                      \
+                 uint8* rgb_buf, int width) {                                  \
+      int n = width & ~MASK;                                                   \
+      if (n > 0) {                                                             \
+        NV12TORGB_SIMD(y_buf, uv_buf, rgb_buf, n);                             \
+      }                                                                        \
       NV12TORGB_C(y_buf + n,                                                   \
                   uv_buf + (n >> UV_SHIFT),                                    \
-                  rgb_buf + n * BPP, width & 7);                               \
+                  rgb_buf + n * BPP, width & MASK);                            \
     }
 
 #ifdef HAS_NV12TOARGBROW_SSSE3
-NV2NY(NV12ToARGBRow_Any_SSSE3, NV12ToARGBRow_Unaligned_SSSE3, NV12ToARGBRow_C,
-      0, 4)
-NV2NY(NV21ToARGBRow_Any_SSSE3, NV21ToARGBRow_Unaligned_SSSE3, NV21ToARGBRow_C,
-      0, 4)
-#endif  // HAS_NV12TOARGBROW_SSSE3
+NV2NY(NV12ToARGBRow_Any_SSSE3, NV12ToARGBRow_SSSE3, NV12ToARGBRow_C, 0, 4, 7)
+NV2NY(NV21ToARGBRow_Any_SSSE3, NV21ToARGBRow_SSSE3, NV21ToARGBRow_C, 0, 4, 7)
+#endif
+#ifdef HAS_NV12TOARGBROW_AVX2
+NV2NY(NV12ToARGBRow_Any_AVX2, NV12ToARGBRow_AVX2, NV12ToARGBRow_C, 0, 4, 15)
+NV2NY(NV21ToARGBRow_Any_AVX2, NV21ToARGBRow_AVX2, NV21ToARGBRow_C, 0, 4, 15)
+#endif
 #ifdef HAS_NV12TOARGBROW_NEON
-NV2NY(NV12ToARGBRow_Any_NEON, NV12ToARGBRow_NEON, NV12ToARGBRow_C, 0, 4)
-NV2NY(NV21ToARGBRow_Any_NEON, NV21ToARGBRow_NEON, NV21ToARGBRow_C, 0, 4)
-#endif  // HAS_NV12TOARGBROW_NEON
+NV2NY(NV12ToARGBRow_Any_NEON, NV12ToARGBRow_NEON, NV12ToARGBRow_C, 0, 4, 7)
+NV2NY(NV21ToARGBRow_Any_NEON, NV21ToARGBRow_NEON, NV21ToARGBRow_C, 0, 4, 7)
+#endif
 #ifdef HAS_NV12TORGB565ROW_SSSE3
 NV2NY(NV12ToRGB565Row_Any_SSSE3, NV12ToRGB565Row_SSSE3, NV12ToRGB565Row_C,
-      0, 2)
+      0, 2, 7)
 NV2NY(NV21ToRGB565Row_Any_SSSE3, NV21ToRGB565Row_SSSE3, NV21ToRGB565Row_C,
-      0, 2)
-#endif  // HAS_NV12TORGB565ROW_SSSE3
+      0, 2, 7)
+#endif
+#ifdef HAS_NV12TORGB565ROW_AVX2
+NV2NY(NV12ToRGB565Row_Any_AVX2, NV12ToRGB565Row_AVX2, NV12ToRGB565Row_C,
+      0, 2, 15)
+NV2NY(NV21ToRGB565Row_Any_AVX2, NV21ToRGB565Row_AVX2, NV21ToRGB565Row_C,
+      0, 2, 15)
+#endif
 #ifdef HAS_NV12TORGB565ROW_NEON
-NV2NY(NV12ToRGB565Row_Any_NEON, NV12ToRGB565Row_NEON, NV12ToRGB565Row_C, 0, 2)
-NV2NY(NV21ToRGB565Row_Any_NEON, NV21ToRGB565Row_NEON, NV21ToRGB565Row_C, 0, 2)
-#endif  // HAS_NV12TORGB565ROW_NEON
+NV2NY(NV12ToRGB565Row_Any_NEON, NV12ToRGB565Row_NEON, NV12ToRGB565Row_C,
+      0, 2, 7)
+NV2NY(NV21ToRGB565Row_Any_NEON, NV21ToRGB565Row_NEON, NV21ToRGB565Row_C,
+      0, 2, 7)
+#endif
 #undef NVANY
 
-#define RGBANY(NAMEANY, ARGBTORGB_SIMD, ARGBTORGB_C, MASK, SBPP, BPP)          \
-    void NAMEANY(const uint8* src,                                             \
-                 uint8* dst,                                                   \
-                 int width) {                                                  \
+#define RGBANY(NAMEANY, ARGBTORGB_SIMD, ARGBTORGB_C, SBPP, BPP, MASK)          \
+    void NAMEANY(const uint8* src, uint8* dst, int width) {                    \
       int n = width & ~MASK;                                                   \
-      ARGBTORGB_SIMD(src, dst, n);                                             \
+      if (n > 0) {                                                             \
+        ARGBTORGB_SIMD(src, dst, n);                                           \
+      }                                                                        \
       ARGBTORGB_C(src + n * SBPP, dst + n * BPP, width & MASK);                \
     }
 
 #if defined(HAS_ARGBTORGB24ROW_SSSE3)
 RGBANY(ARGBToRGB24Row_Any_SSSE3, ARGBToRGB24Row_SSSE3, ARGBToRGB24Row_C,
-       15, 4, 3)
+       4, 3, 15)
 RGBANY(ARGBToRAWRow_Any_SSSE3, ARGBToRAWRow_SSSE3, ARGBToRAWRow_C,
-       15, 4, 3)
+       4, 3, 15)
 RGBANY(ARGBToRGB565Row_Any_SSE2, ARGBToRGB565Row_SSE2, ARGBToRGB565Row_C,
-       3, 4, 2)
+       4, 2, 3)
 RGBANY(ARGBToARGB1555Row_Any_SSE2, ARGBToARGB1555Row_SSE2, ARGBToARGB1555Row_C,
-       3, 4, 2)
+       4, 2, 3)
 RGBANY(ARGBToARGB4444Row_Any_SSE2, ARGBToARGB4444Row_SSE2, ARGBToARGB4444Row_C,
-       3, 4, 2)
+       4, 2, 3)
+#endif
+#if defined(HAS_ARGBTOARGB4444ROW_AVX2)
+RGBANY(ARGBToRGB565Row_Any_AVX2, ARGBToRGB565Row_AVX2, ARGBToRGB565Row_C,
+       4, 2, 7)
+RGBANY(ARGBToARGB1555Row_Any_AVX2, ARGBToARGB1555Row_AVX2, ARGBToARGB1555Row_C,
+       4, 2, 7)
+RGBANY(ARGBToARGB4444Row_Any_AVX2, ARGBToARGB4444Row_AVX2, ARGBToARGB4444Row_C,
+       4, 2, 7)
 #endif
+
 #if defined(HAS_I400TOARGBROW_SSE2)
-RGBANY(I400ToARGBRow_Any_SSE2, I400ToARGBRow_Unaligned_SSE2, I400ToARGBRow_C,
-       7, 1, 4)
+RGBANY(I400ToARGBRow_Any_SSE2, I400ToARGBRow_SSE2, I400ToARGBRow_C, 1, 4, 7)
 #endif
 #if defined(HAS_YTOARGBROW_SSE2)
-RGBANY(YToARGBRow_Any_SSE2, YToARGBRow_SSE2, YToARGBRow_C,
-       7, 1, 4)
-RGBANY(YUY2ToARGBRow_Any_SSSE3, YUY2ToARGBRow_Unaligned_SSSE3, YUY2ToARGBRow_C,
-       15, 2, 4)
-RGBANY(UYVYToARGBRow_Any_SSSE3, UYVYToARGBRow_Unaligned_SSSE3, UYVYToARGBRow_C,
-       15, 2, 4)
-// These require alignment on ARGB, so C is used for remainder.
+RGBANY(YToARGBRow_Any_SSE2, YToARGBRow_SSE2, YToARGBRow_C, 1, 4, 7)
+#endif
+#if defined(HAS_YTOARGBROW_AVX2)
+RGBANY(YToARGBRow_Any_AVX2, YToARGBRow_AVX2, YToARGBRow_C, 1, 4, 15)
+#endif
+#if defined(HAS_YUY2TOARGBROW_SSSE3)
+RGBANY(YUY2ToARGBRow_Any_SSSE3, YUY2ToARGBRow_SSSE3, YUY2ToARGBRow_C, 2, 4, 15)
+RGBANY(UYVYToARGBRow_Any_SSSE3, UYVYToARGBRow_SSSE3, UYVYToARGBRow_C, 2, 4, 15)
 RGBANY(RGB24ToARGBRow_Any_SSSE3, RGB24ToARGBRow_SSSE3, RGB24ToARGBRow_C,
-       15, 3, 4)
-RGBANY(RAWToARGBRow_Any_SSSE3, RAWToARGBRow_SSSE3, RAWToARGBRow_C,
-       15, 3, 4)
+       3, 4, 15)
+RGBANY(RAWToARGBRow_Any_SSSE3, RAWToARGBRow_SSSE3, RAWToARGBRow_C, 3, 4, 15)
 RGBANY(RGB565ToARGBRow_Any_SSE2, RGB565ToARGBRow_SSE2, RGB565ToARGBRow_C,
-       7, 2, 4)
+       2, 4, 7)
 RGBANY(ARGB1555ToARGBRow_Any_SSE2, ARGB1555ToARGBRow_SSE2, ARGB1555ToARGBRow_C,
-       7, 2, 4)
+       2, 4, 7)
 RGBANY(ARGB4444ToARGBRow_Any_SSE2, ARGB4444ToARGBRow_SSE2, ARGB4444ToARGBRow_C,
-       7, 2, 4)
+       2, 4, 7)
+#endif
+#if defined(HAS_YUY2TOARGBROW_AVX2)
+RGBANY(YUY2ToARGBRow_Any_AVX2, YUY2ToARGBRow_AVX2, YUY2ToARGBRow_C, 2, 4, 31)
+RGBANY(UYVYToARGBRow_Any_AVX2, UYVYToARGBRow_AVX2, UYVYToARGBRow_C, 2, 4, 31)
 #endif
 #if defined(HAS_ARGBTORGB24ROW_NEON)
-RGBANY(ARGBToRGB24Row_Any_NEON, ARGBToRGB24Row_NEON, ARGBToRGB24Row_C, 7, 4, 3)
-RGBANY(ARGBToRAWRow_Any_NEON, ARGBToRAWRow_NEON, ARGBToRAWRow_C, 7, 4, 3)
+RGBANY(ARGBToRGB24Row_Any_NEON, ARGBToRGB24Row_NEON, ARGBToRGB24Row_C, 4, 3, 7)
+RGBANY(ARGBToRAWRow_Any_NEON, ARGBToRAWRow_NEON, ARGBToRAWRow_C, 4, 3, 7)
 RGBANY(ARGBToRGB565Row_Any_NEON, ARGBToRGB565Row_NEON, ARGBToRGB565Row_C,
-       7, 4, 2)
+       4, 2, 7)
 RGBANY(ARGBToARGB1555Row_Any_NEON, ARGBToARGB1555Row_NEON, ARGBToARGB1555Row_C,
-       7, 4, 2)
+       4, 2, 7)
 RGBANY(ARGBToARGB4444Row_Any_NEON, ARGBToARGB4444Row_NEON, ARGBToARGB4444Row_C,
-       7, 4, 2)
-RGBANY(I400ToARGBRow_Any_NEON, I400ToARGBRow_NEON, I400ToARGBRow_C,
-       7, 1, 4)
-RGBANY(YToARGBRow_Any_NEON, YToARGBRow_NEON, YToARGBRow_C,
-       7, 1, 4)
-RGBANY(YUY2ToARGBRow_Any_NEON, YUY2ToARGBRow_NEON, YUY2ToARGBRow_C,
-       7, 2, 4)
-RGBANY(UYVYToARGBRow_Any_NEON, UYVYToARGBRow_NEON, UYVYToARGBRow_C,
-       7, 2, 4)
+       4, 2, 7)
+RGBANY(I400ToARGBRow_Any_NEON, I400ToARGBRow_NEON, I400ToARGBRow_C, 1, 4, 7)
+RGBANY(YToARGBRow_Any_NEON, YToARGBRow_NEON, YToARGBRow_C, 1, 4, 7)
+RGBANY(YUY2ToARGBRow_Any_NEON, YUY2ToARGBRow_NEON, YUY2ToARGBRow_C, 2, 4, 7)
+RGBANY(UYVYToARGBRow_Any_NEON, UYVYToARGBRow_NEON, UYVYToARGBRow_C, 2, 4, 7)
 #endif
 #undef RGBANY
 
 // ARGB to Bayer does multiple of 4 pixels, SSSE3 aligned src, unaligned dst.
-#define BAYERANY(NAMEANY, ARGBTORGB_SIMD, ARGBTORGB_C, MASK, SBPP, BPP)        \
-    void NAMEANY(const uint8* src,                                             \
-                 uint8* dst, uint32 selector,                                  \
-                 int width) {                                                  \
+#define BAYERANY(NAMEANY, ARGBTORGB_SIMD, ARGBTORGB_C, SBPP, BPP, MASK)        \
+    void NAMEANY(const uint8* src, uint8* dst, uint32 selector, int width) {   \
       int n = width & ~MASK;                                                   \
-      ARGBTORGB_SIMD(src, dst, selector, n);                                   \
+      if (n > 0) {                                                             \
+        ARGBTORGB_SIMD(src, dst, selector, n);                                 \
+      }                                                                        \
       ARGBTORGB_C(src + n * SBPP, dst + n * BPP, selector, width & MASK);      \
     }
 
-#if defined(HAS_ARGBTOBAYERROW_SSSE3)
-BAYERANY(ARGBToBayerRow_Any_SSSE3, ARGBToBayerRow_SSSE3, ARGBToBayerRow_C,
-         7, 4, 1)
-#endif
-#if defined(HAS_ARGBTOBAYERROW_NEON)
-BAYERANY(ARGBToBayerRow_Any_NEON, ARGBToBayerRow_NEON, ARGBToBayerRow_C,
-         7, 4, 1)
-#endif
 #if defined(HAS_ARGBTOBAYERGGROW_SSE2)
 BAYERANY(ARGBToBayerGGRow_Any_SSE2, ARGBToBayerGGRow_SSE2, ARGBToBayerGGRow_C,
-         7, 4, 1)
+         4, 1, 7)
 #endif
 #if defined(HAS_ARGBTOBAYERGGROW_NEON)
 BAYERANY(ARGBToBayerGGRow_Any_NEON, ARGBToBayerGGRow_NEON, ARGBToBayerGGRow_C,
-         7, 4, 1)
+         4, 1, 7)
 #endif
 
 #undef BAYERANY
 
-// RGB/YUV to Y does multiple of 16 with SIMD and last 16 with SIMD.
-#define YANY(NAMEANY, ARGBTOY_SIMD, SBPP, BPP, NUM)                            \
+#define YANY(NAMEANY, ARGBTOY_SIMD, ARGBTOY_C, SBPP, BPP, MASK)                \
     void NAMEANY(const uint8* src_argb, uint8* dst_y, int width) {             \
-      ARGBTOY_SIMD(src_argb, dst_y, width - NUM);                              \
-      ARGBTOY_SIMD(src_argb + (width - NUM) * SBPP,                            \
-                   dst_y + (width - NUM) * BPP, NUM);                          \
+      int n = width & ~MASK;                                                   \
+      if (n > 0) {                                                             \
+        ARGBTOY_SIMD(src_argb, dst_y, n);                                      \
+      }                                                                        \
+      ARGBTOY_C(src_argb + n * SBPP,                                           \
+                dst_y  + n * BPP, width & MASK);                               \
     }
-
 #ifdef HAS_ARGBTOYROW_AVX2
-YANY(ARGBToYRow_Any_AVX2, ARGBToYRow_AVX2, 4, 1, 32)
-YANY(ARGBToYJRow_Any_AVX2, ARGBToYJRow_AVX2, 4, 1, 32)
-YANY(YUY2ToYRow_Any_AVX2, YUY2ToYRow_AVX2, 2, 1, 32)
-YANY(UYVYToYRow_Any_AVX2, UYVYToYRow_AVX2, 2, 1, 32)
+YANY(ARGBToYRow_Any_AVX2, ARGBToYRow_AVX2, ARGBToYRow_C, 4, 1, 31)
+#endif
+#ifdef HAS_ARGBTOYJROW_AVX2
+YANY(ARGBToYJRow_Any_AVX2, ARGBToYJRow_AVX2, ARGBToYJRow_C, 4, 1, 31)
+#endif
+#ifdef HAS_UYVYTOYROW_AVX2
+YANY(UYVYToYRow_Any_AVX2, UYVYToYRow_AVX2, UYVYToYRow_C, 2, 1, 31)
+#endif
+#ifdef HAS_YUY2TOYROW_AVX2
+YANY(YUY2ToYRow_Any_AVX2, YUY2ToYRow_AVX2, YUY2ToYRow_C, 2, 1, 31)
 #endif
 #ifdef HAS_ARGBTOYROW_SSSE3
-YANY(ARGBToYRow_Any_SSSE3, ARGBToYRow_Unaligned_SSSE3, 4, 1, 16)
+YANY(ARGBToYRow_Any_SSSE3, ARGBToYRow_SSSE3, ARGBToYRow_C, 4, 1, 15)
 #endif
 #ifdef HAS_BGRATOYROW_SSSE3
-YANY(BGRAToYRow_Any_SSSE3, BGRAToYRow_Unaligned_SSSE3, 4, 1, 16)
-YANY(ABGRToYRow_Any_SSSE3, ABGRToYRow_Unaligned_SSSE3, 4, 1, 16)
-YANY(RGBAToYRow_Any_SSSE3, RGBAToYRow_Unaligned_SSSE3, 4, 1, 16)
-YANY(YUY2ToYRow_Any_SSE2, YUY2ToYRow_Unaligned_SSE2, 2, 1, 16)
-YANY(UYVYToYRow_Any_SSE2, UYVYToYRow_Unaligned_SSE2, 2, 1, 16)
+YANY(BGRAToYRow_Any_SSSE3, BGRAToYRow_SSSE3, BGRAToYRow_C, 4, 1, 15)
+YANY(ABGRToYRow_Any_SSSE3, ABGRToYRow_SSSE3, ABGRToYRow_C, 4, 1, 15)
+YANY(RGBAToYRow_Any_SSSE3, RGBAToYRow_SSSE3, RGBAToYRow_C, 4, 1, 15)
+YANY(YUY2ToYRow_Any_SSE2, YUY2ToYRow_SSE2, YUY2ToYRow_C, 2, 1, 15)
+YANY(UYVYToYRow_Any_SSE2, UYVYToYRow_SSE2, UYVYToYRow_C, 2, 1, 15)
 #endif
 #ifdef HAS_ARGBTOYJROW_SSSE3
-YANY(ARGBToYJRow_Any_SSSE3, ARGBToYJRow_Unaligned_SSSE3, 4, 1, 16)
+YANY(ARGBToYJRow_Any_SSSE3, ARGBToYJRow_SSSE3, ARGBToYJRow_C, 4, 1, 15)
 #endif
 #ifdef HAS_ARGBTOYROW_NEON
-YANY(ARGBToYRow_Any_NEON, ARGBToYRow_NEON, 4, 1, 8)
-YANY(ARGBToYJRow_Any_NEON, ARGBToYJRow_NEON, 4, 1, 8)
-YANY(BGRAToYRow_Any_NEON, BGRAToYRow_NEON, 4, 1, 8)
-YANY(ABGRToYRow_Any_NEON, ABGRToYRow_NEON, 4, 1, 8)
-YANY(RGBAToYRow_Any_NEON, RGBAToYRow_NEON, 4, 1, 8)
-YANY(RGB24ToYRow_Any_NEON, RGB24ToYRow_NEON, 3, 1, 8)
-YANY(RAWToYRow_Any_NEON, RAWToYRow_NEON, 3, 1, 8)
-YANY(RGB565ToYRow_Any_NEON, RGB565ToYRow_NEON, 2, 1, 8)
-YANY(ARGB1555ToYRow_Any_NEON, ARGB1555ToYRow_NEON, 2, 1, 8)
-YANY(ARGB4444ToYRow_Any_NEON, ARGB4444ToYRow_NEON, 2, 1, 8)
+YANY(ARGBToYRow_Any_NEON, ARGBToYRow_NEON, ARGBToYRow_C, 4, 1, 7)
+#endif
+#ifdef HAS_ARGBTOYJROW_NEON
+YANY(ARGBToYJRow_Any_NEON, ARGBToYJRow_NEON, ARGBToYJRow_C, 4, 1, 7)
+#endif
+#ifdef HAS_BGRATOYROW_NEON
+YANY(BGRAToYRow_Any_NEON, BGRAToYRow_NEON, BGRAToYRow_C, 4, 1, 7)
+#endif
+#ifdef HAS_ABGRTOYROW_NEON
+YANY(ABGRToYRow_Any_NEON, ABGRToYRow_NEON, ABGRToYRow_C, 4, 1, 7)
+#endif
+#ifdef HAS_RGBATOYROW_NEON
+YANY(RGBAToYRow_Any_NEON, RGBAToYRow_NEON, RGBAToYRow_C, 4, 1, 7)
+#endif
+#ifdef HAS_RGB24TOYROW_NEON
+YANY(RGB24ToYRow_Any_NEON, RGB24ToYRow_NEON, RGB24ToYRow_C, 3, 1, 7)
+#endif
+#ifdef HAS_RAWTOYROW_NEON
+YANY(RAWToYRow_Any_NEON, RAWToYRow_NEON, RAWToYRow_C, 3, 1, 7)
+#endif
+#ifdef HAS_RGB565TOYROW_NEON
+YANY(RGB565ToYRow_Any_NEON, RGB565ToYRow_NEON, RGB565ToYRow_C, 2, 1, 7)
+#endif
+#ifdef HAS_ARGB1555TOYROW_NEON
+YANY(ARGB1555ToYRow_Any_NEON, ARGB1555ToYRow_NEON, ARGB1555ToYRow_C, 2, 1, 7)
+#endif
+#ifdef HAS_ARGB4444TOYROW_NEON
+YANY(ARGB4444ToYRow_Any_NEON, ARGB4444ToYRow_NEON, ARGB4444ToYRow_C, 2, 1, 7)
 #endif
 #ifdef HAS_YUY2TOYROW_NEON
-YANY(YUY2ToYRow_Any_NEON, YUY2ToYRow_NEON, 2, 1, 16)
+YANY(YUY2ToYRow_Any_NEON, YUY2ToYRow_NEON, YUY2ToYRow_C, 2, 1, 15)
 #endif
 #ifdef HAS_UYVYTOYROW_NEON
-YANY(UYVYToYRow_Any_NEON, UYVYToYRow_NEON, 2, 1, 16)
+YANY(UYVYToYRow_Any_NEON, UYVYToYRow_NEON, UYVYToYRow_C, 2, 1, 15)
 #endif
 #ifdef HAS_RGB24TOARGBROW_NEON
-YANY(RGB24ToARGBRow_Any_NEON, RGB24ToARGBRow_NEON, 3, 4, 8)
+YANY(RGB24ToARGBRow_Any_NEON, RGB24ToARGBRow_NEON, RGB24ToARGBRow_C, 3, 4, 7)
 #endif
 #ifdef HAS_RAWTOARGBROW_NEON
-YANY(RAWToARGBRow_Any_NEON, RAWToARGBRow_NEON, 3, 4, 8)
+YANY(RAWToARGBRow_Any_NEON, RAWToARGBRow_NEON, RAWToARGBRow_C, 3, 4, 7)
 #endif
 #ifdef HAS_RGB565TOARGBROW_NEON
-YANY(RGB565ToARGBRow_Any_NEON, RGB565ToARGBRow_NEON, 2, 4, 8)
+YANY(RGB565ToARGBRow_Any_NEON, RGB565ToARGBRow_NEON, RGB565ToARGBRow_C, 2, 4, 7)
 #endif
 #ifdef HAS_ARGB1555TOARGBROW_NEON
-YANY(ARGB1555ToARGBRow_Any_NEON, ARGB1555ToARGBRow_NEON, 2, 4, 8)
+YANY(ARGB1555ToARGBRow_Any_NEON, ARGB1555ToARGBRow_NEON, ARGB1555ToARGBRow_C,
+     2, 4, 7)
 #endif
 #ifdef HAS_ARGB4444TOARGBROW_NEON
-YANY(ARGB4444ToARGBRow_Any_NEON, ARGB4444ToARGBRow_NEON, 2, 4, 8)
+YANY(ARGB4444ToARGBRow_Any_NEON, ARGB4444ToARGBRow_NEON, ARGB4444ToARGBRow_C,
+     2, 4, 7)
 #endif
-#undef YANY
-
-#define YANY(NAMEANY, ARGBTOY_SIMD, ARGBTOY_C, SBPP, BPP, MASK)                \
-    void NAMEANY(const uint8* src_argb, uint8* dst_y, int width) {             \
-      int n = width & ~MASK;                                                   \
-      ARGBTOY_SIMD(src_argb, dst_y, n);                                        \
-      ARGBTOY_C(src_argb + n * SBPP,                                           \
-                dst_y  + n * BPP, width & MASK);                               \
-    }
-
-// Attenuate is destructive so last16 method can not be used due to overlap.
 #ifdef HAS_ARGBATTENUATEROW_SSSE3
 YANY(ARGBAttenuateRow_Any_SSSE3, ARGBAttenuateRow_SSSE3, ARGBAttenuateRow_C,
      4, 4, 3)
@@ -318,7 +364,9 @@ YANY(ARGBAttenuateRow_Any_NEON, ARGBAttenuateRow_NEON, ARGBAttenuateRow_C,
     void NAMEANY(const uint8* src_argb, int src_stride_argb,                   \
                  uint8* dst_u, uint8* dst_v, int width) {                      \
       int n = width & ~MASK;                                                   \
-      ANYTOUV_SIMD(src_argb, src_stride_argb, dst_u, dst_v, n);                \
+      if (n > 0) {                                                             \
+        ANYTOUV_SIMD(src_argb, src_stride_argb, dst_u, dst_v, n);              \
+      }                                                                        \
       ANYTOUV_C(src_argb  + n * BPP, src_stride_argb,                          \
                 dst_u + (n >> 1),                                              \
                 dst_v + (n >> 1),                                              \
@@ -327,29 +375,50 @@ YANY(ARGBAttenuateRow_Any_NEON, ARGBAttenuateRow_NEON, ARGBAttenuateRow_C,
 
 #ifdef HAS_ARGBTOUVROW_AVX2
 UVANY(ARGBToUVRow_Any_AVX2, ARGBToUVRow_AVX2, ARGBToUVRow_C, 4, 31)
+#endif
+#ifdef HAS_ARGBTOUVROW_SSSE3
+UVANY(ARGBToUVRow_Any_SSSE3, ARGBToUVRow_SSSE3, ARGBToUVRow_C, 4, 15)
+UVANY(ARGBToUVJRow_Any_SSSE3, ARGBToUVJRow_SSSE3, ARGBToUVJRow_C, 4, 15)
+UVANY(BGRAToUVRow_Any_SSSE3, BGRAToUVRow_SSSE3, BGRAToUVRow_C, 4, 15)
+UVANY(ABGRToUVRow_Any_SSSE3, ABGRToUVRow_SSSE3, ABGRToUVRow_C, 4, 15)
+UVANY(RGBAToUVRow_Any_SSSE3, RGBAToUVRow_SSSE3, RGBAToUVRow_C, 4, 15)
+#endif
+#ifdef HAS_YUY2TOUVROW_AVX2
 UVANY(YUY2ToUVRow_Any_AVX2, YUY2ToUVRow_AVX2, YUY2ToUVRow_C, 2, 31)
 UVANY(UYVYToUVRow_Any_AVX2, UYVYToUVRow_AVX2, UYVYToUVRow_C, 2, 31)
 #endif
-#ifdef HAS_ARGBTOUVROW_SSSE3
-UVANY(ARGBToUVRow_Any_SSSE3, ARGBToUVRow_Unaligned_SSSE3, ARGBToUVRow_C, 4, 15)
-UVANY(ARGBToUVJRow_Any_SSSE3, ARGBToUVJRow_Unaligned_SSSE3, ARGBToUVJRow_C,
-      4, 15)
-UVANY(BGRAToUVRow_Any_SSSE3, BGRAToUVRow_Unaligned_SSSE3, BGRAToUVRow_C, 4, 15)
-UVANY(ABGRToUVRow_Any_SSSE3, ABGRToUVRow_Unaligned_SSSE3, ABGRToUVRow_C, 4, 15)
-UVANY(RGBAToUVRow_Any_SSSE3, RGBAToUVRow_Unaligned_SSSE3, RGBAToUVRow_C, 4, 15)
-UVANY(YUY2ToUVRow_Any_SSE2, YUY2ToUVRow_Unaligned_SSE2, YUY2ToUVRow_C, 2, 15)
-UVANY(UYVYToUVRow_Any_SSE2, UYVYToUVRow_Unaligned_SSE2, UYVYToUVRow_C, 2, 15)
+#ifdef HAS_YUY2TOUVROW_SSE2
+UVANY(YUY2ToUVRow_Any_SSE2, YUY2ToUVRow_SSE2, YUY2ToUVRow_C, 2, 15)
+UVANY(UYVYToUVRow_Any_SSE2, UYVYToUVRow_SSE2, UYVYToUVRow_C, 2, 15)
 #endif
 #ifdef HAS_ARGBTOUVROW_NEON
 UVANY(ARGBToUVRow_Any_NEON, ARGBToUVRow_NEON, ARGBToUVRow_C, 4, 15)
+#endif
+#ifdef HAS_ARGBTOUVJROW_NEON
 UVANY(ARGBToUVJRow_Any_NEON, ARGBToUVJRow_NEON, ARGBToUVJRow_C, 4, 15)
+#endif
+#ifdef HAS_BGRATOUVROW_NEON
 UVANY(BGRAToUVRow_Any_NEON, BGRAToUVRow_NEON, BGRAToUVRow_C, 4, 15)
+#endif
+#ifdef HAS_ABGRTOUVROW_NEON
 UVANY(ABGRToUVRow_Any_NEON, ABGRToUVRow_NEON, ABGRToUVRow_C, 4, 15)
+#endif
+#ifdef HAS_RGBATOUVROW_NEON
 UVANY(RGBAToUVRow_Any_NEON, RGBAToUVRow_NEON, RGBAToUVRow_C, 4, 15)
+#endif
+#ifdef HAS_RGB24TOUVROW_NEON
 UVANY(RGB24ToUVRow_Any_NEON, RGB24ToUVRow_NEON, RGB24ToUVRow_C, 3, 15)
+#endif
+#ifdef HAS_RAWTOUVROW_NEON
 UVANY(RAWToUVRow_Any_NEON, RAWToUVRow_NEON, RAWToUVRow_C, 3, 15)
+#endif
+#ifdef HAS_RGB565TOUVROW_NEON
 UVANY(RGB565ToUVRow_Any_NEON, RGB565ToUVRow_NEON, RGB565ToUVRow_C, 2, 15)
+#endif
+#ifdef HAS_ARGB1555TOUVROW_NEON
 UVANY(ARGB1555ToUVRow_Any_NEON, ARGB1555ToUVRow_NEON, ARGB1555ToUVRow_C, 2, 15)
+#endif
+#ifdef HAS_ARGB4444TOUVROW_NEON
 UVANY(ARGB4444ToUVRow_Any_NEON, ARGB4444ToUVRow_NEON, ARGB4444ToUVRow_C, 2, 15)
 #endif
 #ifdef HAS_YUY2TOUVROW_NEON
@@ -360,11 +429,12 @@ UVANY(UYVYToUVRow_Any_NEON, UYVYToUVRow_NEON, UYVYToUVRow_C, 2, 15)
 #endif
 #undef UVANY
 
-#define UV422ANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, BPP, MASK, SHIFT)           \
-    void NAMEANY(const uint8* src_uv,                                          \
-                 uint8* dst_u, uint8* dst_v, int width) {                      \
+#define UV422ANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, BPP, SHIFT, MASK)           \
+    void NAMEANY(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int width) { \
       int n = width & ~MASK;                                                   \
-      ANYTOUV_SIMD(src_uv, dst_u, dst_v, n);                                   \
+      if (n > 0) {                                                             \
+        ANYTOUV_SIMD(src_uv, dst_u, dst_v, n);                                 \
+      }                                                                        \
       ANYTOUV_C(src_uv  + n * BPP,                                             \
                 dst_u + (n >> SHIFT),                                          \
                 dst_v + (n >> SHIFT),                                          \
@@ -372,42 +442,45 @@ UVANY(UYVYToUVRow_Any_NEON, UYVYToUVRow_NEON, UYVYToUVRow_C, 2, 15)
     }
 
 #ifdef HAS_ARGBTOUV444ROW_SSSE3
-UV422ANY(ARGBToUV444Row_Any_SSSE3, ARGBToUV444Row_Unaligned_SSSE3,
-         ARGBToUV444Row_C, 4, 15, 0)
+UV422ANY(ARGBToUV444Row_Any_SSSE3, ARGBToUV444Row_SSSE3,
+         ARGBToUV444Row_C, 4, 0, 15)
 #endif
 #ifdef HAS_YUY2TOUV422ROW_AVX2
 UV422ANY(YUY2ToUV422Row_Any_AVX2, YUY2ToUV422Row_AVX2,
-         YUY2ToUV422Row_C, 2, 31, 1)
+         YUY2ToUV422Row_C, 2, 1, 31)
 UV422ANY(UYVYToUV422Row_Any_AVX2, UYVYToUV422Row_AVX2,
-         UYVYToUV422Row_C, 2, 31, 1)
+         UYVYToUV422Row_C, 2, 1, 31)
 #endif
-#ifdef HAS_ARGBTOUVROW_SSSE3
-UV422ANY(ARGBToUV422Row_Any_SSSE3, ARGBToUV422Row_Unaligned_SSSE3,
-         ARGBToUV422Row_C, 4, 15, 1)
-UV422ANY(YUY2ToUV422Row_Any_SSE2, YUY2ToUV422Row_Unaligned_SSE2,
-         YUY2ToUV422Row_C, 2, 15, 1)
-UV422ANY(UYVYToUV422Row_Any_SSE2, UYVYToUV422Row_Unaligned_SSE2,
-         UYVYToUV422Row_C, 2, 15, 1)
+#ifdef HAS_ARGBTOUV422ROW_SSSE3
+UV422ANY(ARGBToUV422Row_Any_SSSE3, ARGBToUV422Row_SSSE3,
+         ARGBToUV422Row_C, 4, 1, 15)
+#endif
+#ifdef HAS_YUY2TOUV422ROW_SSE2
+UV422ANY(YUY2ToUV422Row_Any_SSE2, YUY2ToUV422Row_SSE2,
+         YUY2ToUV422Row_C, 2, 1, 15)
+UV422ANY(UYVYToUV422Row_Any_SSE2, UYVYToUV422Row_SSE2,
+         UYVYToUV422Row_C, 2, 1, 15)
 #endif
 #ifdef HAS_YUY2TOUV422ROW_NEON
 UV422ANY(ARGBToUV444Row_Any_NEON, ARGBToUV444Row_NEON,
-         ARGBToUV444Row_C, 4, 7, 0)
+         ARGBToUV444Row_C, 4, 0, 7)
 UV422ANY(ARGBToUV422Row_Any_NEON, ARGBToUV422Row_NEON,
-         ARGBToUV422Row_C, 4, 15, 1)
+         ARGBToUV422Row_C, 4, 1, 15)
 UV422ANY(ARGBToUV411Row_Any_NEON, ARGBToUV411Row_NEON,
-         ARGBToUV411Row_C, 4, 31, 2)
+         ARGBToUV411Row_C, 4, 2, 31)
 UV422ANY(YUY2ToUV422Row_Any_NEON, YUY2ToUV422Row_NEON,
-         YUY2ToUV422Row_C, 2, 15, 1)
+         YUY2ToUV422Row_C, 2, 1, 15)
 UV422ANY(UYVYToUV422Row_Any_NEON, UYVYToUV422Row_NEON,
-         UYVYToUV422Row_C, 2, 15, 1)
+         UYVYToUV422Row_C, 2, 1, 15)
 #endif
 #undef UV422ANY
 
 #define SPLITUVROWANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, MASK)                  \
-    void NAMEANY(const uint8* src_uv,                                          \
-                 uint8* dst_u, uint8* dst_v, int width) {                      \
+    void NAMEANY(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int width) { \
       int n = width & ~MASK;                                                   \
-      ANYTOUV_SIMD(src_uv, dst_u, dst_v, n);                                   \
+      if (n > 0) {                                                             \
+        ANYTOUV_SIMD(src_uv, dst_u, dst_v, n);                                 \
+      }                                                                        \
       ANYTOUV_C(src_uv + n * 2,                                                \
                 dst_u + n,                                                     \
                 dst_v + n,                                                     \
@@ -415,7 +488,7 @@ UV422ANY(UYVYToUV422Row_Any_NEON, UYVYToUV422Row_NEON,
     }
 
 #ifdef HAS_SPLITUVROW_SSE2
-SPLITUVROWANY(SplitUVRow_Any_SSE2, SplitUVRow_Unaligned_SSE2, SplitUVRow_C, 15)
+SPLITUVROWANY(SplitUVRow_Any_SSE2, SplitUVRow_SSE2, SplitUVRow_C, 15)
 #endif
 #ifdef HAS_SPLITUVROW_AVX2
 SPLITUVROWANY(SplitUVRow_Any_AVX2, SplitUVRow_AVX2, SplitUVRow_C, 31)
@@ -424,7 +497,7 @@ SPLITUVROWANY(SplitUVRow_Any_AVX2, SplitUVRow_AVX2, SplitUVRow_C, 31)
 SPLITUVROWANY(SplitUVRow_Any_NEON, SplitUVRow_NEON, SplitUVRow_C, 15)
 #endif
 #ifdef HAS_SPLITUVROW_MIPS_DSPR2
-SPLITUVROWANY(SplitUVRow_Any_MIPS_DSPR2, SplitUVRow_Unaligned_MIPS_DSPR2,
+SPLITUVROWANY(SplitUVRow_Any_MIPS_DSPR2, SplitUVRow_MIPS_DSPR2,
               SplitUVRow_C, 15)
 #endif
 #undef SPLITUVROWANY
@@ -433,7 +506,9 @@ SPLITUVROWANY(SplitUVRow_Any_MIPS_DSPR2, SplitUVRow_Unaligned_MIPS_DSPR2,
     void NAMEANY(const uint8* src_u, const uint8* src_v,                       \
                  uint8* dst_uv, int width) {                                   \
       int n = width & ~MASK;                                                   \
-      ANYTOUV_SIMD(src_u, src_v, dst_uv, n);                                   \
+      if (n > 0) {                                                             \
+        ANYTOUV_SIMD(src_u, src_v, dst_uv, n);                                 \
+      }                                                                        \
       ANYTOUV_C(src_u + n,                                                     \
                 src_v + n,                                                     \
                 dst_uv + n * 2,                                                \
@@ -441,7 +516,7 @@ SPLITUVROWANY(SplitUVRow_Any_MIPS_DSPR2, SplitUVRow_Unaligned_MIPS_DSPR2,
     }
 
 #ifdef HAS_MERGEUVROW_SSE2
-MERGEUVROW_ANY(MergeUVRow_Any_SSE2, MergeUVRow_Unaligned_SSE2, MergeUVRow_C, 15)
+MERGEUVROW_ANY(MergeUVRow_Any_SSE2, MergeUVRow_SSE2, MergeUVRow_C, 15)
 #endif
 #ifdef HAS_MERGEUVROW_AVX2
 MERGEUVROW_ANY(MergeUVRow_Any_AVX2, MergeUVRow_AVX2, MergeUVRow_C, 31)
@@ -455,7 +530,9 @@ MERGEUVROW_ANY(MergeUVRow_Any_NEON, MergeUVRow_NEON, MergeUVRow_C, 15)
     void NAMEANY(const uint8* src_argb0, const uint8* src_argb1,               \
                  uint8* dst_argb, int width) {                                 \
       int n = width & ~MASK;                                                   \
-      ARGBMATH_SIMD(src_argb0, src_argb1, dst_argb, n);                        \
+      if (n > 0) {                                                             \
+        ARGBMATH_SIMD(src_argb0, src_argb1, dst_argb, n);                      \
+      }                                                                        \
       ARGBMATH_C(src_argb0 + n * 4,                                            \
                  src_argb1 + n * 4,                                            \
                  dst_argb + n * 4,                                             \
@@ -502,7 +579,9 @@ MATHROW_ANY(ARGBSubtractRow_Any_NEON, ARGBSubtractRow_NEON, ARGBSubtractRow_C,
     void NAMEANY(const uint8* src_argb, uint8* dst_argb,                       \
                  const uint8* shuffler, int width) {                           \
       int n = width & ~MASK;                                                   \
-      ARGBTOY_SIMD(src_argb, dst_argb, shuffler, n);                           \
+      if (n > 0) {                                                             \
+        ARGBTOY_SIMD(src_argb, dst_argb, shuffler, n);                         \
+      }                                                                        \
       ARGBTOY_C(src_argb + n * SBPP,                                           \
                 dst_argb  + n * BPP, shuffler, width & MASK);                  \
     }
@@ -512,7 +591,7 @@ YANY(ARGBShuffleRow_Any_SSE2, ARGBShuffleRow_SSE2,
      ARGBShuffleRow_C, 4, 4, 3)
 #endif
 #ifdef HAS_ARGBSHUFFLEROW_SSSE3
-YANY(ARGBShuffleRow_Any_SSSE3, ARGBShuffleRow_Unaligned_SSSE3,
+YANY(ARGBShuffleRow_Any_SSSE3, ARGBShuffleRow_SSSE3,
      ARGBShuffleRow_C, 4, 4, 7)
 #endif
 #ifdef HAS_ARGBSHUFFLEROW_AVX2
@@ -531,35 +610,107 @@ YANY(ARGBShuffleRow_Any_NEON, ARGBShuffleRow_NEON,
                  ptrdiff_t src_stride_ptr, int width,                          \
                  int source_y_fraction) {                                      \
       int n = width & ~MASK;                                                   \
-      TERP_SIMD(dst_ptr, src_ptr, src_stride_ptr,                              \
-                n, source_y_fraction);                                         \
+      if (n > 0) {                                                             \
+        TERP_SIMD(dst_ptr, src_ptr, src_stride_ptr, n, source_y_fraction);     \
+      }                                                                        \
       TERP_C(dst_ptr + n * BPP,                                                \
              src_ptr + n * SBPP, src_stride_ptr,                               \
              width & MASK, source_y_fraction);                                 \
     }
 
 #ifdef HAS_INTERPOLATEROW_AVX2
-NANY(InterpolateRow_Any_AVX2, InterpolateRow_AVX2,
-     InterpolateRow_C, 1, 1, 32)
+NANY(InterpolateRow_Any_AVX2, InterpolateRow_AVX2, InterpolateRow_C, 1, 1, 31)
 #endif
 #ifdef HAS_INTERPOLATEROW_SSSE3
-NANY(InterpolateRow_Any_SSSE3, InterpolateRow_Unaligned_SSSE3,
-     InterpolateRow_C, 1, 1, 15)
+NANY(InterpolateRow_Any_SSSE3, InterpolateRow_SSSE3, InterpolateRow_C, 1, 1, 15)
 #endif
 #ifdef HAS_INTERPOLATEROW_SSE2
-NANY(InterpolateRow_Any_SSE2, InterpolateRow_Unaligned_SSE2,
-     InterpolateRow_C, 1, 1, 15)
+NANY(InterpolateRow_Any_SSE2, InterpolateRow_SSE2, InterpolateRow_C, 1, 1, 15)
 #endif
 #ifdef HAS_INTERPOLATEROW_NEON
-NANY(InterpolateRow_Any_NEON, InterpolateRow_NEON,
-     InterpolateRow_C, 1, 1, 15)
+NANY(InterpolateRow_Any_NEON, InterpolateRow_NEON, InterpolateRow_C, 1, 1, 15)
 #endif
 #ifdef HAS_INTERPOLATEROW_MIPS_DSPR2
-NANY(InterpolateRow_Any_MIPS_DSPR2, InterpolateRow_MIPS_DSPR2,
-     InterpolateRow_C, 1, 1, 3)
+NANY(InterpolateRow_Any_MIPS_DSPR2, InterpolateRow_MIPS_DSPR2, InterpolateRow_C,
+     1, 1, 3)
 #endif
 #undef NANY
 
+#define MANY(NAMEANY, MIRROR_SIMD, MIRROR_C, BPP, MASK)                        \
+    void NAMEANY(const uint8* src_y, uint8* dst_y, int width) {                \
+      int n = width & ~MASK;                                                   \
+      int r = width & MASK;                                                    \
+      if (n > 0) {                                                             \
+        MIRROR_SIMD(src_y, dst_y + r * BPP, n);                                \
+      }                                                                        \
+      MIRROR_C(src_y + n * BPP, dst_y, r);                                     \
+    }
+
+#ifdef HAS_MIRRORROW_AVX2
+MANY(MirrorRow_Any_AVX2, MirrorRow_AVX2, MirrorRow_C, 1, 31)
+#endif
+#ifdef HAS_MIRRORROW_SSSE3
+MANY(MirrorRow_Any_SSSE3, MirrorRow_SSSE3, MirrorRow_C, 1, 15)
+#endif
+#ifdef HAS_MIRRORROW_SSE2
+MANY(MirrorRow_Any_SSE2, MirrorRow_SSE2, MirrorRow_C, 1, 15)
+#endif
+#ifdef HAS_MIRRORROW_NEON
+MANY(MirrorRow_Any_NEON, MirrorRow_NEON, MirrorRow_C, 1, 15)
+#endif
+#ifdef HAS_ARGBMIRRORROW_AVX2
+MANY(ARGBMirrorRow_Any_AVX2, ARGBMirrorRow_AVX2, ARGBMirrorRow_C, 4, 7)
+#endif
+#ifdef HAS_ARGBMIRRORROW_SSE2
+MANY(ARGBMirrorRow_Any_SSE2, ARGBMirrorRow_SSE2, ARGBMirrorRow_C, 4, 3)
+#endif
+#ifdef HAS_ARGBMIRRORROW_NEON
+MANY(ARGBMirrorRow_Any_NEON, ARGBMirrorRow_NEON, ARGBMirrorRow_C, 4, 3)
+#endif
+#undef MANY
+
+#define MANY(NAMEANY, COPY_SIMD, COPY_C, BPP, MASK)                            \
+    void NAMEANY(const uint8* src_y, uint8* dst_y, int width) {                \
+      int n = width & ~MASK;                                                   \
+      int r = width & MASK;                                                    \
+      if (n > 0) {                                                             \
+        COPY_SIMD(src_y, dst_y, n);                                            \
+      }                                                                        \
+      COPY_C(src_y + n * BPP, dst_y + n * BPP, r);                             \
+    }
+
+#ifdef HAS_COPYROW_AVX
+MANY(CopyRow_Any_AVX, CopyRow_AVX, CopyRow_C, 1, 63)
+#endif
+#ifdef HAS_COPYROW_SSE2
+MANY(CopyRow_Any_SSE2, CopyRow_SSE2, CopyRow_C, 1, 31)
+#endif
+#ifdef HAS_COPYROW_NEON
+MANY(CopyRow_Any_NEON, CopyRow_NEON, CopyRow_C, 1, 31)
+#endif
+#undef MANY
+
+#define SETANY(NAMEANY, SET_SIMD, SET_C, T, BPP, MASK)                         \
+    void NAMEANY(uint8* dst_y, T v8, int width) {                              \
+      int n = width & ~MASK;                                                   \
+      int r = width & MASK;                                                    \
+      if (n > 0) {                                                             \
+        SET_SIMD(dst_y, v8, n);                                                \
+      }                                                                        \
+      SET_C(dst_y + n * BPP, v8, r);                                           \
+    }
+
+#ifdef HAS_SETROW_X86
+SETANY(SetRow_Any_X86, SetRow_X86, SetRow_ERMS, uint8, 1, 3)
+#endif
+#ifdef HAS_SETROW_NEON
+SETANY(SetRow_Any_NEON, SetRow_NEON, SetRow_C, uint8, 1, 15)
+#endif
+#ifdef HAS_ARGBSETROW_NEON
+SETANY(ARGBSetRow_Any_NEON, ARGBSetRow_NEON, ARGBSetRow_C, uint32, 4, 3)
+#endif
+#undef SETANY
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
diff --git a/third_party/libyuv/source/row_common.cc b/third_party/libyuv/source/row_common.cc
index fa2b752a2..e0e2bf426 100644
--- a/third_party/libyuv/source/row_common.cc
+++ b/third_party/libyuv/source/row_common.cc
@@ -199,6 +199,32 @@ void ARGBToRGB565Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
   }
 }
 
+void ARGBToRGB565DitherRow_C(const uint8* src_argb, uint8* dst_rgb,
+                             const uint8* dither8x8, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    int dither0 = dither8x8[x & 7] - 128;
+    int dither1 = dither8x8[(x & 7) + 1] - 128;
+    uint8 b0 = Clamp(src_argb[0] + dither0) >> 3;
+    uint8 g0 = Clamp(src_argb[1] + dither0) >> 2;
+    uint8 r0 = Clamp(src_argb[2] + dither0) >> 3;
+    uint8 b1 = Clamp(src_argb[4] + dither1) >> 3;
+    uint8 g1 = Clamp(src_argb[5] + dither1) >> 2;
+    uint8 r1 = Clamp(src_argb[6] + dither1) >> 3;
+    WRITEWORD(dst_rgb, b0 | (g0 << 5) | (r0 << 11) |
+              (b1 << 16) | (g1 << 21) | (r1 << 27));
+    dst_rgb += 4;
+    src_argb += 8;
+  }
+  if (width & 1) {
+    int dither0 = dither8x8[(width - 1) & 7] - 128;
+    uint8 b0 = Clamp(src_argb[0] + dither0) >> 3;
+    uint8 g0 = Clamp(src_argb[1] + dither0) >> 2;
+    uint8 r0 = Clamp(src_argb[2] + dither0) >> 3;
+    *(uint16*)(dst_rgb) = b0 | (g0 << 5) | (r0 << 11);
+  }
+}
+
 void ARGBToARGB1555Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
@@ -385,6 +411,28 @@ void NAME ## ToUVJRow_C(const uint8* src_rgb0, int src_stride_rgb,             \
 MAKEROWYJ(ARGB, 2, 1, 0, 4)
 #undef MAKEROWYJ
 
+void ARGBToUVJ422Row_C(const uint8* src_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 ab = (src_argb[0] + src_argb[4]) >> 1;
+    uint8 ag = (src_argb[1] + src_argb[5]) >> 1;
+    uint8 ar = (src_argb[2] + src_argb[6]) >> 1;
+    dst_u[0] = RGBToUJ(ar, ag, ab);
+    dst_v[0] = RGBToVJ(ar, ag, ab);
+    src_argb += 8;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if (width & 1) {
+    uint8 ab = src_argb[0];
+    uint8 ag = src_argb[1];
+    uint8 ar = src_argb[2];
+    dst_u[0] = RGBToUJ(ar, ag, ab);
+    dst_v[0] = RGBToVJ(ar, ag, ab);
+  }
+}
+
 void RGB565ToYRow_C(const uint8* src_rgb565, uint8* dst_y, int width) {
   int x;
   for (x = 0; x < width; ++x) {
@@ -938,33 +986,52 @@ void I400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int width) {
   }
 }
 
-// C reference code that mimics the YUV assembly.
-
-#define YG 74 /* (int8)(1.164 * 64 + 0.5) */
-
-#define UB 127 /* min(63,(int8)(2.018 * 64)) */
-#define UG -25 /* (int8)(-0.391 * 64 - 0.5) */
-#define UR 0
+// YUV to RGB conversion constants.
+// Y contribution to R,G,B.  Scale and bias.
+// TODO(fbarchard): Consider moving constants into a common header.
+#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */
+#define YGB 1160 /* 1.164 * 64 * 16 - adjusted for even error distribution */
 
-#define VB 0
-#define VG -52 /* (int8)(-0.813 * 64 - 0.5) */
-#define VR 102 /* (int8)(1.596 * 64 + 0.5) */
+// U and V contributions to R,G,B.
+#define UB -128 /* -min(128, round(2.018 * 64)) */
+#define UG 25 /* -round(-0.391 * 64) */
+#define VG 52 /* -round(-0.813 * 64) */
+#define VR -102 /* -round(1.596 * 64) */
 
-// Bias
-#define BB UB * 128 + VB * 128
-#define BG UG * 128 + VG * 128
-#define BR UR * 128 + VR * 128
+// Bias values to subtract 16 from Y and 128 from U and V.
+#define BB (UB * 128            - YGB)
+#define BG (UG * 128 + VG * 128 - YGB)
+#define BR            (VR * 128 - YGB)
 
+// C reference code that mimics the YUV assembly.
 static __inline void YuvPixel(uint8 y, uint8 u, uint8 v,
                               uint8* b, uint8* g, uint8* r) {
-  int32 y1 = ((int32)(y) - 16) * YG;
-  *b = Clamp((int32)((u * UB + v * VB) - (BB) + y1) >> 6);
-  *g = Clamp((int32)((u * UG + v * VG) - (BG) + y1) >> 6);
-  *r = Clamp((int32)((u * UR + v * VR) - (BR) + y1) >> 6);
+  uint32 y1 = (uint32)(y * 0x0101 * YG) >> 16;
+  *b = Clamp((int32)(BB - (         u * UB) + y1) >> 6);
+  *g = Clamp((int32)(BG - (v * VG + u * UG) + y1) >> 6);
+  *r = Clamp((int32)(BR - (v * VR         ) + y1) >> 6);
 }
 
+// C reference code that mimics the YUV assembly.
+static __inline void YPixel(uint8 y, uint8* b, uint8* g, uint8* r) {
+  uint32 y1 = (uint32)(y * 0x0101 * YG) >> 16;
+  *b = Clamp((int32)(y1 - YGB) >> 6);
+  *g = Clamp((int32)(y1 - YGB) >> 6);
+  *r = Clamp((int32)(y1 - YGB) >> 6);
+}
+
+#undef YG
+#undef YGB
+#undef UB
+#undef UG
+#undef VG
+#undef VR
+#undef BB
+#undef BG
+#undef BR
+
 #if !defined(LIBYUV_DISABLE_NEON) && \
-    (defined(__ARM_NEON__) || defined(LIBYUV_NEON))
+    (defined(__ARM_NEON__) || defined(__aarch64__) || defined(LIBYUV_NEON))
 // C mimic assembly.
 // TODO(fbarchard): Remove subsampling from Neon.
 void I444ToARGBRow_C(const uint8* src_y,
@@ -1008,6 +1075,7 @@ void I444ToARGBRow_C(const uint8* src_y,
   }
 }
 #endif
+
 // Also used for 420
 void I422ToARGBRow_C(const uint8* src_y,
                      const uint8* src_u,
@@ -1034,6 +1102,59 @@ void I422ToARGBRow_C(const uint8* src_y,
   }
 }
 
+// C reference code that mimics the YUV assembly.
+// *  R = Y                + 1.40200 * Cr
+// *  G = Y - 0.34414 * Cb - 0.71414 * Cr
+// *  B = Y + 1.77200 * Cb
+
+#define YGJ 64 /* (int8)round(1.000 * 64) */
+
+#define UBJ 113 /* (int8)round(1.772 * 64) */
+#define UGJ -22 /* (int8)round(-0.34414 * 64) */
+#define URJ 0
+
+#define VBJ 0
+#define VGJ -46 /* (int8)round(-0.71414 * 64) */
+#define VRJ 90 /* (int8)round(1.402 * 64) */
+
+// Bias
+#define BBJ (UBJ * 128 + VBJ * 128)
+#define BGJ (UGJ * 128 + VGJ * 128)
+#define BRJ (URJ * 128 + VRJ * 128)
+
+static __inline void YuvJPixel(uint8 y, uint8 u, uint8 v,
+                              uint8* b, uint8* g, uint8* r) {
+  uint32 y1 = (uint32)(y * YGJ);
+  *b = Clamp((int32)(u * UBJ + v * VBJ + y1 - BBJ) >> 6);
+  *g = Clamp((int32)(u * UGJ + v * VGJ + y1 - BGJ) >> 6);
+  *r = Clamp((int32)(u * URJ + v * VRJ + y1 - BRJ) >> 6);
+}
+
+void J422ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvJPixel(src_y[0], src_u[0], src_v[0],
+              rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvJPixel(src_y[1], src_u[0], src_v[0],
+              rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvJPixel(src_y[0], src_u[0], src_v[0],
+              rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
 void I422ToRGB24Row_C(const uint8* src_y,
                       const uint8* src_u,
                       const uint8* src_v,
@@ -1470,18 +1591,15 @@ void I422ToRGBARow_C(const uint8* src_y,
 void YToARGBRow_C(const uint8* src_y, uint8* rgb_buf, int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
-    YuvPixel(src_y[0], 128, 128,
-             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    YPixel(src_y[0], rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
     rgb_buf[3] = 255;
-    YuvPixel(src_y[1], 128, 128,
-             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    YPixel(src_y[1], rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
     rgb_buf[7] = 255;
     src_y += 2;
     rgb_buf += 8;  // Advance 2 pixels.
   }
   if (width & 1) {
-    YuvPixel(src_y[0], 128, 128,
-             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    YPixel(src_y[0], rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
     rgb_buf[3] = 255;
   }
 }
@@ -1569,28 +1687,15 @@ void CopyRow_16_C(const uint16* src, uint16* dst, int count) {
   memcpy(dst, src, count * 2);
 }
 
-void SetRow_C(uint8* dst, uint32 v8, int count) {
-#ifdef _MSC_VER
-  // VC will generate rep stosb.
-  int x;
-  for (x = 0; x < count; ++x) {
-    dst[x] = v8;
-  }
-#else
-  memset(dst, v8, count);
-#endif
+void SetRow_C(uint8* dst, uint8 v8, int width) {
+  memset(dst, v8, width);
 }
 
-void ARGBSetRows_C(uint8* dst, uint32 v32, int width,
-                 int dst_stride, int height) {
-  int y;
-  for (y = 0; y < height; ++y) {
-    uint32* d = (uint32*)(dst);
-    int x;
-    for (x = 0; x < width; ++x) {
-      d[x] = v32;
-    }
-    dst += dst_stride;
+void ARGBSetRow_C(uint8* dst_argb, uint32 v32, int width) {
+  uint32* d = (uint32*)(dst_argb);
+  int x;
+  for (x = 0; x < width; ++x) {
+    d[x] = v32;
   }
 }
 
@@ -1885,17 +1990,17 @@ void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride,
   }
 }
 
-// Blend 2 rows into 1 for conversions such as I422ToI420.
-void HalfRow_C(const uint8* src_uv, int src_uv_stride,
-               uint8* dst_uv, int pix) {
+// Blend 2 rows into 1.
+static void HalfRow_C(const uint8* src_uv, int src_uv_stride,
+                      uint8* dst_uv, int pix) {
   int x;
   for (x = 0; x < pix; ++x) {
     dst_uv[x] = (src_uv[x] + src_uv[src_uv_stride + x] + 1) >> 1;
   }
 }
 
-void HalfRow_16_C(const uint16* src_uv, int src_uv_stride,
-                  uint16* dst_uv, int pix) {
+static void HalfRow_16_C(const uint16* src_uv, int src_uv_stride,
+                         uint16* dst_uv, int pix) {
   int x;
   for (x = 0; x < pix; ++x) {
     dst_uv[x] = (src_uv[x] + src_uv[src_uv_stride + x] + 1) >> 1;
@@ -1957,24 +2062,6 @@ void InterpolateRow_16_C(uint16* dst_ptr, const uint16* src_ptr,
   }
 }
 
-// Select 2 channels from ARGB on alternating pixels.  e.g.  BGBGBGBG
-void ARGBToBayerRow_C(const uint8* src_argb,
-                      uint8* dst_bayer, uint32 selector, int pix) {
-  int index0 = selector & 0xff;
-  int index1 = (selector >> 8) & 0xff;
-  // Copy a row of Bayer.
-  int x;
-  for (x = 0; x < pix - 1; x += 2) {
-    dst_bayer[0] = src_argb[index0];
-    dst_bayer[1] = src_argb[index1];
-    src_argb += 8;
-    dst_bayer += 2;
-  }
-  if (pix & 1) {
-    dst_bayer[0] = src_argb[index0];
-  }
-}
-
 // Select G channel from ARGB.  e.g.  GGGGGGGG
 void ARGBToBayerGGRow_C(const uint8* src_argb,
                         uint8* dst_bayer, uint32 selector, int pix) {
@@ -2061,122 +2148,272 @@ void I422ToUYVYRow_C(const uint8* src_y,
   }
 }
 
-#if !defined(LIBYUV_DISABLE_X86) && defined(HAS_I422TOARGBROW_SSSE3)
+// Maximum temporary width for wrappers to process at a time, in pixels.
+#define MAXTWIDTH 2048
+
+#if !defined(_MSC_VER) && defined(HAS_I422TORGB565ROW_SSSE3)
 // row_win.cc has asm version, but GCC uses 2 step wrapper.
-#if !defined(_MSC_VER) && (defined(__x86_64__) || defined(__i386__))
 void I422ToRGB565Row_SSSE3(const uint8* src_y,
                            const uint8* src_u,
                            const uint8* src_v,
-                           uint8* rgb_buf,
+                           uint8* dst_rgb565,
                            int width) {
-  // Allocate a row of ARGB.
-  align_buffer_64(row, width * 4);
-  I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, width);
-  ARGBToRGB565Row_SSE2(row, rgb_buf, width);
-  free_aligned_buffer_64(row);
+  SIMD_ALIGNED(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, twidth);
+    ARGBToRGB565Row_SSE2(row, dst_rgb565, twidth);
+    src_y += twidth;
+    src_u += twidth / 2;
+    src_v += twidth / 2;
+    dst_rgb565 += twidth * 2;
+    width -= twidth;
+  }
 }
-#endif  // !defined(_MSC_VER) && (defined(__x86_64__) || defined(__i386__))
+#endif
 
-#if defined(_M_IX86) || defined(__x86_64__) || defined(__i386__)
+#if defined(HAS_I422TOARGB1555ROW_SSSE3)
 void I422ToARGB1555Row_SSSE3(const uint8* src_y,
                              const uint8* src_u,
                              const uint8* src_v,
-                             uint8* rgb_buf,
+                             uint8* dst_argb1555,
                              int width) {
-  // Allocate a row of ARGB.
-  align_buffer_64(row, width * 4);
-  I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, width);
-  ARGBToARGB1555Row_SSE2(row, rgb_buf, width);
-  free_aligned_buffer_64(row);
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, twidth);
+    ARGBToARGB1555Row_SSE2(row, dst_argb1555, twidth);
+    src_y += twidth;
+    src_u += twidth / 2;
+    src_v += twidth / 2;
+    dst_argb1555 += twidth * 2;
+    width -= twidth;
+  }
 }
+#endif
 
+#if defined(HAS_I422TOARGB4444ROW_SSSE3)
 void I422ToARGB4444Row_SSSE3(const uint8* src_y,
                              const uint8* src_u,
                              const uint8* src_v,
-                             uint8* rgb_buf,
+                             uint8* dst_argb4444,
                              int width) {
-  // Allocate a row of ARGB.
-  align_buffer_64(row, width * 4);
-  I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, width);
-  ARGBToARGB4444Row_SSE2(row, rgb_buf, width);
-  free_aligned_buffer_64(row);
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, twidth);
+    ARGBToARGB4444Row_SSE2(row, dst_argb4444, twidth);
+    src_y += twidth;
+    src_u += twidth / 2;
+    src_v += twidth / 2;
+    dst_argb4444 += twidth * 2;
+    width -= twidth;
+  }
 }
+#endif
 
-void NV12ToRGB565Row_SSSE3(const uint8* src_y,
-                           const uint8* src_uv,
-                           uint8* dst_rgb565,
-                           int width) {
-  // Allocate a row of ARGB.
-  align_buffer_64(row, width * 4);
-  NV12ToARGBRow_SSSE3(src_y, src_uv, row, width);
-  ARGBToRGB565Row_SSE2(row, dst_rgb565, width);
-  free_aligned_buffer_64(row);
+#if defined(HAS_NV12TORGB565ROW_SSSE3)
+void NV12ToRGB565Row_SSSE3(const uint8* src_y, const uint8* src_uv,
+                           uint8* dst_rgb565, int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    NV12ToARGBRow_SSSE3(src_y, src_uv, row, twidth);
+    ARGBToRGB565Row_SSE2(row, dst_rgb565, twidth);
+    src_y += twidth;
+    src_uv += twidth;
+    dst_rgb565 += twidth * 2;
+    width -= twidth;
+  }
 }
+#endif
 
-void NV21ToRGB565Row_SSSE3(const uint8* src_y,
-                           const uint8* src_vu,
-                           uint8* dst_rgb565,
-                           int width) {
-  // Allocate a row of ARGB.
-  align_buffer_64(row, width * 4);
-  NV21ToARGBRow_SSSE3(src_y, src_vu, row, width);
-  ARGBToRGB565Row_SSE2(row, dst_rgb565, width);
-  free_aligned_buffer_64(row);
+#if defined(HAS_NV21TORGB565ROW_SSSE3)
+void NV21ToRGB565Row_SSSE3(const uint8* src_y, const uint8* src_vu,
+                           uint8* dst_rgb565, int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    NV21ToARGBRow_SSSE3(src_y, src_vu, row, twidth);
+    ARGBToRGB565Row_SSE2(row, dst_rgb565, twidth);
+    src_y += twidth;
+    src_vu += twidth;
+    dst_rgb565 += twidth * 2;
+    width -= twidth;
+  }
 }
+#endif
 
-void YUY2ToARGBRow_SSSE3(const uint8* src_yuy2,
-                         uint8* dst_argb,
-                         int width) {
-  // Allocate a rows of yuv.
-  align_buffer_64(row_y, ((width + 63) & ~63) * 2);
-  uint8* row_u = row_y + ((width + 63) & ~63);
-  uint8* row_v = row_u + ((width + 63) & ~63) / 2;
-  YUY2ToUV422Row_SSE2(src_yuy2, row_u, row_v, width);
-  YUY2ToYRow_SSE2(src_yuy2, row_y, width);
-  I422ToARGBRow_SSSE3(row_y, row_u, row_v, dst_argb, width);
-  free_aligned_buffer_64(row_y);
-}
-
-void YUY2ToARGBRow_Unaligned_SSSE3(const uint8* src_yuy2,
-                                   uint8* dst_argb,
-                                   int width) {
-  // Allocate a rows of yuv.
-  align_buffer_64(row_y, ((width + 63) & ~63) * 2);
-  uint8* row_u = row_y + ((width + 63) & ~63);
-  uint8* row_v = row_u + ((width + 63) & ~63) / 2;
-  YUY2ToUV422Row_Unaligned_SSE2(src_yuy2, row_u, row_v, width);
-  YUY2ToYRow_Unaligned_SSE2(src_yuy2, row_y, width);
-  I422ToARGBRow_Unaligned_SSSE3(row_y, row_u, row_v, dst_argb, width);
-  free_aligned_buffer_64(row_y);
-}
-
-void UYVYToARGBRow_SSSE3(const uint8* src_uyvy,
-                         uint8* dst_argb,
-                         int width) {
-  // Allocate a rows of yuv.
-  align_buffer_64(row_y, ((width + 63) & ~63) * 2);
-  uint8* row_u = row_y + ((width + 63) & ~63);
-  uint8* row_v = row_u + ((width + 63) & ~63) / 2;
-  UYVYToUV422Row_SSE2(src_uyvy, row_u, row_v, width);
-  UYVYToYRow_SSE2(src_uyvy, row_y, width);
-  I422ToARGBRow_SSSE3(row_y, row_u, row_v, dst_argb, width);
-  free_aligned_buffer_64(row_y);
-}
-
-void UYVYToARGBRow_Unaligned_SSSE3(const uint8* src_uyvy,
-                                   uint8* dst_argb,
-                                   int width) {
-  // Allocate a rows of yuv.
-  align_buffer_64(row_y, ((width + 63) & ~63) * 2);
-  uint8* row_u = row_y + ((width + 63) & ~63);
-  uint8* row_v = row_u + ((width + 63) & ~63) / 2;
-  UYVYToUV422Row_Unaligned_SSE2(src_uyvy, row_u, row_v, width);
-  UYVYToYRow_Unaligned_SSE2(src_uyvy, row_y, width);
-  I422ToARGBRow_Unaligned_SSSE3(row_y, row_u, row_v, dst_argb, width);
-  free_aligned_buffer_64(row_y);
-}
-
-#endif  // defined(_M_IX86) || defined(__x86_64__) || defined(__i386__)
+#if defined(HAS_YUY2TOARGBROW_SSSE3)
+void YUY2ToARGBRow_SSSE3(const uint8* src_yuy2, uint8* dst_argb, int width) {
+  // Row buffers for intermediate YUV pixels.
+  SIMD_ALIGNED(uint8 row_y[MAXTWIDTH]);
+  SIMD_ALIGNED(uint8 row_u[MAXTWIDTH / 2]);
+  SIMD_ALIGNED(uint8 row_v[MAXTWIDTH / 2]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    YUY2ToUV422Row_SSE2(src_yuy2, row_u, row_v, twidth);
+    YUY2ToYRow_SSE2(src_yuy2, row_y, twidth);
+    I422ToARGBRow_SSSE3(row_y, row_u, row_v, dst_argb, twidth);
+    src_yuy2 += twidth * 2;
+    dst_argb += twidth * 4;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_UYVYTOARGBROW_SSSE3)
+void UYVYToARGBRow_SSSE3(const uint8* src_uyvy, uint8* dst_argb, int width) {
+  // Row buffers for intermediate YUV pixels.
+  SIMD_ALIGNED(uint8 row_y[MAXTWIDTH]);
+  SIMD_ALIGNED(uint8 row_u[MAXTWIDTH / 2]);
+  SIMD_ALIGNED(uint8 row_v[MAXTWIDTH / 2]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    UYVYToUV422Row_SSE2(src_uyvy, row_u, row_v, twidth);
+    UYVYToYRow_SSE2(src_uyvy, row_y, twidth);
+    I422ToARGBRow_SSSE3(row_y, row_u, row_v, dst_argb, twidth);
+    src_uyvy += twidth * 2;
+    dst_argb += twidth * 4;
+    width -= twidth;
+  }
+}
+#endif  // !defined(LIBYUV_DISABLE_X86)
+
+#if defined(HAS_I422TORGB565ROW_AVX2)
+void I422ToRGB565Row_AVX2(const uint8* src_y,
+                          const uint8* src_u,
+                          const uint8* src_v,
+                          uint8* dst_rgb565,
+                          int width) {
+  SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    I422ToARGBRow_AVX2(src_y, src_u, src_v, row, twidth);
+    ARGBToRGB565Row_AVX2(row, dst_rgb565, twidth);
+    src_y += twidth;
+    src_u += twidth / 2;
+    src_v += twidth / 2;
+    dst_rgb565 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_I422TOARGB1555ROW_AVX2)
+void I422ToARGB1555Row_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb1555,
+                            int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    I422ToARGBRow_AVX2(src_y, src_u, src_v, row, twidth);
+    ARGBToARGB1555Row_AVX2(row, dst_argb1555, twidth);
+    src_y += twidth;
+    src_u += twidth / 2;
+    src_v += twidth / 2;
+    dst_argb1555 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_I422TOARGB4444ROW_AVX2)
+void I422ToARGB4444Row_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb4444,
+                            int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    I422ToARGBRow_AVX2(src_y, src_u, src_v, row, twidth);
+    ARGBToARGB4444Row_AVX2(row, dst_argb4444, twidth);
+    src_y += twidth;
+    src_u += twidth / 2;
+    src_v += twidth / 2;
+    dst_argb4444 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_NV12TORGB565ROW_AVX2)
+void NV12ToRGB565Row_AVX2(const uint8* src_y, const uint8* src_uv,
+                          uint8* dst_rgb565, int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    NV12ToARGBRow_AVX2(src_y, src_uv, row, twidth);
+    ARGBToRGB565Row_AVX2(row, dst_rgb565, twidth);
+    src_y += twidth;
+    src_uv += twidth;
+    dst_rgb565 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_NV21TORGB565ROW_AVX2)
+void NV21ToRGB565Row_AVX2(const uint8* src_y, const uint8* src_vu,
+                          uint8* dst_rgb565, int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    NV21ToARGBRow_AVX2(src_y, src_vu, row, twidth);
+    ARGBToRGB565Row_AVX2(row, dst_rgb565, twidth);
+    src_y += twidth;
+    src_vu += twidth;
+    dst_rgb565 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_YUY2TOARGBROW_AVX2)
+void YUY2ToARGBRow_AVX2(const uint8* src_yuy2, uint8* dst_argb, int width) {
+  // Row buffers for intermediate YUV pixels.
+  SIMD_ALIGNED32(uint8 row_y[MAXTWIDTH]);
+  SIMD_ALIGNED32(uint8 row_u[MAXTWIDTH / 2]);
+  SIMD_ALIGNED32(uint8 row_v[MAXTWIDTH / 2]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    YUY2ToUV422Row_AVX2(src_yuy2, row_u, row_v, twidth);
+    YUY2ToYRow_AVX2(src_yuy2, row_y, twidth);
+    I422ToARGBRow_AVX2(row_y, row_u, row_v, dst_argb, twidth);
+    src_yuy2 += twidth * 2;
+    dst_argb += twidth * 4;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_UYVYTOARGBROW_AVX2)
+void UYVYToARGBRow_AVX2(const uint8* src_uyvy, uint8* dst_argb, int width) {
+  // Row buffers for intermediate YUV pixels.
+  SIMD_ALIGNED32(uint8 row_y[MAXTWIDTH]);
+  SIMD_ALIGNED32(uint8 row_u[MAXTWIDTH / 2]);
+  SIMD_ALIGNED32(uint8 row_v[MAXTWIDTH / 2]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    UYVYToUV422Row_AVX2(src_uyvy, row_u, row_v, twidth);
+    UYVYToYRow_AVX2(src_uyvy, row_y, twidth);
+    I422ToARGBRow_AVX2(row_y, row_u, row_v, dst_argb, twidth);
+    src_uyvy += twidth * 2;
+    dst_argb += twidth * 4;
+    width -= twidth;
+  }
+}
 #endif  // !defined(LIBYUV_DISABLE_X86)
 
 void ARGBPolynomialRow_C(const uint8* src_argb,
diff --git a/third_party/libyuv/source/row_mips.cc b/third_party/libyuv/source/row_mips.cc
index ae9370c1b..cfc9ffe03 100644
--- a/third_party/libyuv/source/row_mips.cc
+++ b/third_party/libyuv/source/row_mips.cc
@@ -378,7 +378,7 @@ void CopyRow_MIPS(const uint8* src, uint8* dst, int count) {
 // MIPS DSPR2 functions
 #if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips_dsp) && \
     (__mips_dsp_rev >= 2) && \
-    (_MIPS_SIM == _MIPS_SIM_ABI32)
+    (_MIPS_SIM == _MIPS_SIM_ABI32) && (__mips_isa_rev < 6)
 
 void SplitUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
                            int width) {
@@ -447,89 +447,6 @@ void SplitUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
   );
 }
 
-void SplitUVRow_Unaligned_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u,
-                                     uint8* dst_v, int width) {
-  __asm__ __volatile__ (
-    ".set push                                     \n"
-    ".set noreorder                                \n"
-    "srl             $t4, %[width], 4              \n"  // multiplies of 16
-    "blez            $t4, 2f                       \n"
-    " andi           %[width], %[width], 0xf       \n"  // residual
-
-    ".p2align        2                             \n"
-  "1:                                              \n"
-    "addiu           $t4, $t4, -1                  \n"
-    "lwr             $t0, 0(%[src_uv])             \n"
-    "lwl             $t0, 3(%[src_uv])             \n"  // V1 | U1 | V0 | U0
-    "lwr             $t1, 4(%[src_uv])             \n"
-    "lwl             $t1, 7(%[src_uv])             \n"  // V3 | U3 | V2 | U2
-    "lwr             $t2, 8(%[src_uv])             \n"
-    "lwl             $t2, 11(%[src_uv])            \n"  // V5 | U5 | V4 | U4
-    "lwr             $t3, 12(%[src_uv])            \n"
-    "lwl             $t3, 15(%[src_uv])            \n"  // V7 | U7 | V6 | U6
-    "lwr             $t5, 16(%[src_uv])            \n"
-    "lwl             $t5, 19(%[src_uv])            \n"  // V9 | U9 | V8 | U8
-    "lwr             $t6, 20(%[src_uv])            \n"
-    "lwl             $t6, 23(%[src_uv])            \n"  // V11 | U11 | V10 | U10
-    "lwr             $t7, 24(%[src_uv])            \n"
-    "lwl             $t7, 27(%[src_uv])            \n"  // V13 | U13 | V12 | U12
-    "lwr             $t8, 28(%[src_uv])            \n"
-    "lwl             $t8, 31(%[src_uv])            \n"  // V15 | U15 | V14 | U14
-    "precrq.qb.ph    $t9, $t1, $t0                 \n"  // V3 | V2 | V1 | V0
-    "precr.qb.ph     $t0, $t1, $t0                 \n"  // U3 | U2 | U1 | U0
-    "precrq.qb.ph    $t1, $t3, $t2                 \n"  // V7 | V6 | V5 | V4
-    "precr.qb.ph     $t2, $t3, $t2                 \n"  // U7 | U6 | U5 | U4
-    "precrq.qb.ph    $t3, $t6, $t5                 \n"  // V11 | V10 | V9 | V8
-    "precr.qb.ph     $t5, $t6, $t5                 \n"  // U11 | U10 | U9 | U8
-    "precrq.qb.ph    $t6, $t8, $t7                 \n"  // V15 | V14 | V13 | V12
-    "precr.qb.ph     $t7, $t8, $t7                 \n"  // U15 | U14 | U13 | U12
-    "addiu           %[src_uv], %[src_uv], 32      \n"
-    "swr             $t9, 0(%[dst_v])              \n"
-    "swl             $t9, 3(%[dst_v])              \n"
-    "swr             $t0, 0(%[dst_u])              \n"
-    "swl             $t0, 3(%[dst_u])              \n"
-    "swr             $t1, 4(%[dst_v])              \n"
-    "swl             $t1, 7(%[dst_v])              \n"
-    "swr             $t2, 4(%[dst_u])              \n"
-    "swl             $t2, 7(%[dst_u])              \n"
-    "swr             $t3, 8(%[dst_v])              \n"
-    "swl             $t3, 11(%[dst_v])             \n"
-    "swr             $t5, 8(%[dst_u])              \n"
-    "swl             $t5, 11(%[dst_u])             \n"
-    "swr             $t6, 12(%[dst_v])             \n"
-    "swl             $t6, 15(%[dst_v])             \n"
-    "swr             $t7, 12(%[dst_u])             \n"
-    "swl             $t7, 15(%[dst_u])             \n"
-    "addiu           %[dst_u], %[dst_u], 16        \n"
-    "bgtz            $t4, 1b                       \n"
-    " addiu          %[dst_v], %[dst_v], 16        \n"
-
-    "beqz            %[width], 3f                  \n"
-    " nop                                          \n"
-
-  "2:                                              \n"
-    "lbu             $t0, 0(%[src_uv])             \n"
-    "lbu             $t1, 1(%[src_uv])             \n"
-    "addiu           %[src_uv], %[src_uv], 2       \n"
-    "addiu           %[width], %[width], -1        \n"
-    "sb              $t0, 0(%[dst_u])              \n"
-    "sb              $t1, 0(%[dst_v])              \n"
-    "addiu           %[dst_u], %[dst_u], 1         \n"
-    "bgtz            %[width], 2b                  \n"
-    " addiu          %[dst_v], %[dst_v], 1         \n"
-
-  "3:                                              \n"
-    ".set pop                                      \n"
-     : [src_uv] "+r" (src_uv),
-       [width] "+r" (width),
-       [dst_u] "+r" (dst_u),
-       [dst_v] "+r" (dst_v)
-     :
-     : "t0", "t1", "t2", "t3",
-     "t4", "t5", "t6", "t7", "t8", "t9"
-  );
-}
-
 void MirrorRow_MIPS_DSPR2(const uint8* src, uint8* dst, int width) {
   __asm__ __volatile__ (
     ".set push                             \n"
@@ -927,9 +844,9 @@ void I422ToBGRARow_MIPS_DSPR2(const uint8* y_buf,
 }
 
 // Bilinear filter 8x2 -> 8x1
-void InterpolateRows_MIPS_DSPR2(uint8* dst_ptr, const uint8* src_ptr,
-                                ptrdiff_t src_stride, int dst_width,
-                                int source_y_fraction) {
+void InterpolateRow_MIPS_DSPR2(uint8* dst_ptr, const uint8* src_ptr,
+                               ptrdiff_t src_stride, int dst_width,
+                               int source_y_fraction) {
     int y0_fraction = 256 - source_y_fraction;
     const uint8* src_ptr1 = src_ptr + src_stride;
 
diff --git a/third_party/libyuv/source/row_neon.cc b/third_party/libyuv/source/row_neon.cc
index 1392cf5fc..8badc5a9b 100644
--- a/third_party/libyuv/source/row_neon.cc
+++ b/third_party/libyuv/source/row_neon.cc
@@ -16,7 +16,8 @@ extern "C" {
 #endif
 
 // This module is for GCC Neon
-#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__)
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \
+    !defined(__aarch64__)
 
 // Read 8 Y, 4 U and 4 V from 422
 #define READYUV422                                                             \
@@ -92,36 +93,73 @@ extern "C" {
     "vuzp.u8    d2, d3                         \n"                             \
     "vtrn.u32   d2, d3                         \n"
 
+#define YUV422TORGB_SETUP_REG                                                  \
+    "vld1.8     {d24}, [%[kUVToRB]]            \n"                             \
+    "vld1.8     {d25}, [%[kUVToG]]             \n"                             \
+    "vld1.16    {d26[], d27[]}, [%[kUVBiasBGR]]! \n"                           \
+    "vld1.16    {d8[], d9[]}, [%[kUVBiasBGR]]!   \n"                           \
+    "vld1.16    {d28[], d29[]}, [%[kUVBiasBGR]]  \n"                           \
+    "vld1.32    {d30[], d31[]}, [%[kYToRgb]]     \n"
+
 #define YUV422TORGB                                                            \
-    "veor.u8    d2, d26                        \n"/*subtract 128 from u and v*/\
-    "vmull.s8   q8, d2, d24                    \n"/*  u/v B/R component      */\
-    "vmull.s8   q9, d2, d25                    \n"/*  u/v G component        */\
-    "vmov.u8    d1, #0                         \n"/*  split odd/even y apart */\
-    "vtrn.u8    d0, d1                         \n"                             \
-    "vsub.s16   q0, q0, q15                    \n"/*  offset y               */\
-    "vmul.s16   q0, q0, q14                    \n"                             \
+    "vmull.u8   q8, d2, d24                    \n" /* u/v B/R component      */\
+    "vmull.u8   q9, d2, d25                    \n" /* u/v G component        */\
+    "vmovl.u8   q0, d0                         \n" /* Y                      */\
+    "vmovl.s16  q10, d1                        \n"                             \
+    "vmovl.s16  q0, d0                         \n"                             \
+    "vmul.s32   q10, q10, q15                  \n"                             \
+    "vmul.s32   q0, q0, q15                    \n"                             \
+    "vqshrun.s32 d0, q0, #16                   \n"                             \
+    "vqshrun.s32 d1, q10, #16                  \n" /* Y                      */\
     "vadd.s16   d18, d19                       \n"                             \
-    "vqadd.s16  d20, d0, d16                   \n" /* B */                     \
-    "vqadd.s16  d21, d1, d16                   \n"                             \
-    "vqadd.s16  d22, d0, d17                   \n" /* R */                     \
-    "vqadd.s16  d23, d1, d17                   \n"                             \
-    "vqadd.s16  d16, d0, d18                   \n" /* G */                     \
-    "vqadd.s16  d17, d1, d18                   \n"                             \
-    "vqshrun.s16 d0, q10, #6                   \n" /* B */                     \
-    "vqshrun.s16 d1, q11, #6                   \n" /* G */                     \
-    "vqshrun.s16 d2, q8, #6                    \n" /* R */                     \
-    "vmovl.u8   q10, d0                        \n"/*  set up for reinterleave*/\
-    "vmovl.u8   q11, d1                        \n"                             \
-    "vmovl.u8   q8, d2                         \n"                             \
-    "vtrn.u8    d20, d21                       \n"                             \
-    "vtrn.u8    d22, d23                       \n"                             \
-    "vtrn.u8    d16, d17                       \n"                             \
-    "vmov.u8    d21, d16                       \n"
-
-static vec8 kUVToRB  = { 127, 127, 127, 127, 102, 102, 102, 102,
-                         0, 0, 0, 0, 0, 0, 0, 0 };
-static vec8 kUVToG = { -25, -25, -25, -25, -52, -52, -52, -52,
-                       0, 0, 0, 0, 0, 0, 0, 0 };
+    "vshll.u16  q1, d16, #16                   \n" /* Replicate u * UB       */\
+    "vshll.u16  q10, d17, #16                  \n" /* Replicate v * VR       */\
+    "vshll.u16  q3, d18, #16                   \n" /* Replicate (v*VG + u*UG)*/\
+    "vaddw.u16  q1, q1, d16                    \n"                             \
+    "vaddw.u16  q10, q10, d17                  \n"                             \
+    "vaddw.u16  q3, q3, d18                    \n"                             \
+    "vqadd.s16  q8, q0, q13                    \n" /* B */                     \
+    "vqadd.s16  q9, q0, q14                    \n" /* R */                     \
+    "vqadd.s16  q0, q0, q4                     \n" /* G */                     \
+    "vqadd.s16  q8, q8, q1                     \n" /* B */                     \
+    "vqadd.s16  q9, q9, q10                    \n" /* R */                     \
+    "vqsub.s16  q0, q0, q3                     \n" /* G */                     \
+    "vqshrun.s16 d20, q8, #6                   \n" /* B */                     \
+    "vqshrun.s16 d22, q9, #6                   \n" /* R */                     \
+    "vqshrun.s16 d21, q0, #6                   \n" /* G */
+
+// YUV to RGB conversion constants.
+// Y contribution to R,G,B.  Scale and bias.
+#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */
+#define YGB 1160 /* 1.164 * 64 * 16 - adjusted for even error distribution */
+
+// U and V contributions to R,G,B.
+#define UB -128 /* -min(128, round(2.018 * 64)) */
+#define UG 25 /* -round(-0.391 * 64) */
+#define VG 52 /* -round(-0.813 * 64) */
+#define VR -102 /* -round(1.596 * 64) */
+
+// Bias values to subtract 16 from Y and 128 from U and V.
+#define BB (UB * 128            - YGB)
+#define BG (UG * 128 + VG * 128 - YGB)
+#define BR            (VR * 128 - YGB)
+
+static uvec8 kUVToRB  = { 128, 128, 128, 128, 102, 102, 102, 102,
+                          0, 0, 0, 0, 0, 0, 0, 0 };
+static uvec8 kUVToG = { 25, 25, 25, 25, 52, 52, 52, 52,
+                        0, 0, 0, 0, 0, 0, 0, 0 };
+static vec16 kUVBiasBGR = { BB, BG, BR, 0, 0, 0, 0, 0 };
+static vec32 kYToRgb = { 0x0101 * YG, 0, 0, 0 };
+
+#undef YG
+#undef YGB
+#undef UB
+#undef UG
+#undef VG
+#undef VR
+#undef BB
+#undef BG
+#undef BR
 
 void I444ToARGBRow_NEON(const uint8* src_y,
                         const uint8* src_u,
@@ -129,13 +167,7 @@ void I444ToARGBRow_NEON(const uint8* src_y,
                         uint8* dst_argb,
                         int width) {
   asm volatile (
-    MEMACCESS(5)
-    "vld1.8     {d24}, [%5]                    \n"
-    MEMACCESS(6)
-    "vld1.8     {d25}, [%6]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
+    YUV422TORGB_SETUP_REG
     ".p2align   2                              \n"
   "1:                                          \n"
     READYUV444
@@ -150,8 +182,10 @@ void I444ToARGBRow_NEON(const uint8* src_y,
       "+r"(src_v),     // %2
       "+r"(dst_argb),  // %3
       "+r"(width)      // %4
-    : "r"(&kUVToRB),   // %5
-      "r"(&kUVToG)     // %6
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
     : "cc", "memory", "q0", "q1", "q2", "q3",
       "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
   );
@@ -163,13 +197,7 @@ void I422ToARGBRow_NEON(const uint8* src_y,
                         uint8* dst_argb,
                         int width) {
   asm volatile (
-    MEMACCESS(5)
-    "vld1.8     {d24}, [%5]                    \n"
-    MEMACCESS(6)
-    "vld1.8     {d25}, [%6]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
+    YUV422TORGB_SETUP_REG
     ".p2align   2                              \n"
   "1:                                          \n"
     READYUV422
@@ -184,8 +212,10 @@ void I422ToARGBRow_NEON(const uint8* src_y,
       "+r"(src_v),     // %2
       "+r"(dst_argb),  // %3
       "+r"(width)      // %4
-    : "r"(&kUVToRB),   // %5
-      "r"(&kUVToG)     // %6
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
     : "cc", "memory", "q0", "q1", "q2", "q3",
       "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
   );
@@ -197,13 +227,7 @@ void I411ToARGBRow_NEON(const uint8* src_y,
                         uint8* dst_argb,
                         int width) {
   asm volatile (
-    MEMACCESS(5)
-    "vld1.8     {d24}, [%5]                    \n"
-    MEMACCESS(6)
-    "vld1.8     {d25}, [%6]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
+    YUV422TORGB_SETUP_REG
     ".p2align   2                              \n"
   "1:                                          \n"
     READYUV411
@@ -218,8 +242,10 @@ void I411ToARGBRow_NEON(const uint8* src_y,
       "+r"(src_v),     // %2
       "+r"(dst_argb),  // %3
       "+r"(width)      // %4
-    : "r"(&kUVToRB),   // %5
-      "r"(&kUVToG)     // %6
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
     : "cc", "memory", "q0", "q1", "q2", "q3",
       "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
   );
@@ -231,13 +257,7 @@ void I422ToBGRARow_NEON(const uint8* src_y,
                         uint8* dst_bgra,
                         int width) {
   asm volatile (
-    MEMACCESS(5)
-    "vld1.8     {d24}, [%5]                    \n"
-    MEMACCESS(6)
-    "vld1.8     {d25}, [%6]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
+    YUV422TORGB_SETUP_REG
     ".p2align   2                              \n"
   "1:                                          \n"
     READYUV422
@@ -253,8 +273,10 @@ void I422ToBGRARow_NEON(const uint8* src_y,
       "+r"(src_v),     // %2
       "+r"(dst_bgra),  // %3
       "+r"(width)      // %4
-    : "r"(&kUVToRB),   // %5
-      "r"(&kUVToG)     // %6
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
     : "cc", "memory", "q0", "q1", "q2", "q3",
       "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
   );
@@ -266,13 +288,7 @@ void I422ToABGRRow_NEON(const uint8* src_y,
                         uint8* dst_abgr,
                         int width) {
   asm volatile (
-    MEMACCESS(5)
-    "vld1.8     {d24}, [%5]                    \n"
-    MEMACCESS(6)
-    "vld1.8     {d25}, [%6]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
+    YUV422TORGB_SETUP_REG
     ".p2align   2                              \n"
   "1:                                          \n"
     READYUV422
@@ -288,8 +304,10 @@ void I422ToABGRRow_NEON(const uint8* src_y,
       "+r"(src_v),     // %2
       "+r"(dst_abgr),  // %3
       "+r"(width)      // %4
-    : "r"(&kUVToRB),   // %5
-      "r"(&kUVToG)     // %6
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
     : "cc", "memory", "q0", "q1", "q2", "q3",
       "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
   );
@@ -301,13 +319,7 @@ void I422ToRGBARow_NEON(const uint8* src_y,
                         uint8* dst_rgba,
                         int width) {
   asm volatile (
-    MEMACCESS(5)
-    "vld1.8     {d24}, [%5]                    \n"
-    MEMACCESS(6)
-    "vld1.8     {d25}, [%6]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
+    YUV422TORGB_SETUP_REG
     ".p2align   2                              \n"
   "1:                                          \n"
     READYUV422
@@ -322,8 +334,10 @@ void I422ToRGBARow_NEON(const uint8* src_y,
       "+r"(src_v),     // %2
       "+r"(dst_rgba),  // %3
       "+r"(width)      // %4
-    : "r"(&kUVToRB),   // %5
-      "r"(&kUVToG)     // %6
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
     : "cc", "memory", "q0", "q1", "q2", "q3",
       "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
   );
@@ -335,13 +349,7 @@ void I422ToRGB24Row_NEON(const uint8* src_y,
                          uint8* dst_rgb24,
                          int width) {
   asm volatile (
-    MEMACCESS(5)
-    "vld1.8     {d24}, [%5]                    \n"
-    MEMACCESS(6)
-    "vld1.8     {d25}, [%6]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
+    YUV422TORGB_SETUP_REG
     ".p2align   2                              \n"
   "1:                                          \n"
     READYUV422
@@ -355,8 +363,10 @@ void I422ToRGB24Row_NEON(const uint8* src_y,
       "+r"(src_v),      // %2
       "+r"(dst_rgb24),  // %3
       "+r"(width)       // %4
-    : "r"(&kUVToRB),    // %5
-      "r"(&kUVToG)      // %6
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
     : "cc", "memory", "q0", "q1", "q2", "q3",
       "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
   );
@@ -368,13 +378,7 @@ void I422ToRAWRow_NEON(const uint8* src_y,
                        uint8* dst_raw,
                        int width) {
   asm volatile (
-    MEMACCESS(5)
-    "vld1.8     {d24}, [%5]                    \n"
-    MEMACCESS(6)
-    "vld1.8     {d25}, [%6]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
+    YUV422TORGB_SETUP_REG
     ".p2align   2                              \n"
   "1:                                          \n"
     READYUV422
@@ -389,8 +393,10 @@ void I422ToRAWRow_NEON(const uint8* src_y,
       "+r"(src_v),    // %2
       "+r"(dst_raw),  // %3
       "+r"(width)     // %4
-    : "r"(&kUVToRB),  // %5
-      "r"(&kUVToG)    // %6
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
     : "cc", "memory", "q0", "q1", "q2", "q3",
       "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
   );
@@ -414,13 +420,7 @@ void I422ToRGB565Row_NEON(const uint8* src_y,
                           uint8* dst_rgb565,
                           int width) {
   asm volatile (
-    MEMACCESS(5)
-    "vld1.8     {d24}, [%5]                    \n"
-    MEMACCESS(6)
-    "vld1.8     {d25}, [%6]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
+    YUV422TORGB_SETUP_REG
     ".p2align   2                              \n"
   "1:                                          \n"
     READYUV422
@@ -435,8 +435,10 @@ void I422ToRGB565Row_NEON(const uint8* src_y,
       "+r"(src_v),    // %2
       "+r"(dst_rgb565),  // %3
       "+r"(width)     // %4
-    : "r"(&kUVToRB),  // %5
-      "r"(&kUVToG)    // %6
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
     : "cc", "memory", "q0", "q1", "q2", "q3",
       "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
   );
@@ -463,13 +465,7 @@ void I422ToARGB1555Row_NEON(const uint8* src_y,
                             uint8* dst_argb1555,
                             int width) {
   asm volatile (
-    MEMACCESS(5)
-    "vld1.8     {d24}, [%5]                    \n"
-    MEMACCESS(6)
-    "vld1.8     {d25}, [%6]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
+    YUV422TORGB_SETUP_REG
     ".p2align   2                              \n"
   "1:                                          \n"
     READYUV422
@@ -485,8 +481,10 @@ void I422ToARGB1555Row_NEON(const uint8* src_y,
       "+r"(src_v),    // %2
       "+r"(dst_argb1555),  // %3
       "+r"(width)     // %4
-    : "r"(&kUVToRB),  // %5
-      "r"(&kUVToG)    // %6
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
     : "cc", "memory", "q0", "q1", "q2", "q3",
       "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
   );
@@ -507,13 +505,7 @@ void I422ToARGB4444Row_NEON(const uint8* src_y,
                             uint8* dst_argb4444,
                             int width) {
   asm volatile (
-    MEMACCESS(5)
-    "vld1.8     {d24}, [%5]                    \n"
-    MEMACCESS(6)
-    "vld1.8     {d25}, [%6]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
+    YUV422TORGB_SETUP_REG
     "vmov.u8    d4, #0x0f                      \n"  // bits to clear with vbic.
     ".p2align   2                              \n"
   "1:                                          \n"
@@ -530,8 +522,10 @@ void I422ToARGB4444Row_NEON(const uint8* src_y,
       "+r"(src_v),    // %2
       "+r"(dst_argb4444),  // %3
       "+r"(width)     // %4
-    : "r"(&kUVToRB),  // %5
-      "r"(&kUVToG)    // %6
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
     : "cc", "memory", "q0", "q1", "q2", "q3",
       "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
   );
@@ -541,13 +535,7 @@ void YToARGBRow_NEON(const uint8* src_y,
                      uint8* dst_argb,
                      int width) {
   asm volatile (
-    MEMACCESS(3)
-    "vld1.8     {d24}, [%3]                    \n"
-    MEMACCESS(4)
-    "vld1.8     {d25}, [%4]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
+    YUV422TORGB_SETUP_REG
     ".p2align   2                              \n"
   "1:                                          \n"
     READYUV400
@@ -560,8 +548,10 @@ void YToARGBRow_NEON(const uint8* src_y,
     : "+r"(src_y),     // %0
       "+r"(dst_argb),  // %1
       "+r"(width)      // %2
-    : "r"(&kUVToRB),   // %3
-      "r"(&kUVToG)     // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %3
+      [kUVToG]"r"(&kUVToG),     // %4
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
     : "cc", "memory", "q0", "q1", "q2", "q3",
       "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
   );
@@ -595,13 +585,7 @@ void NV12ToARGBRow_NEON(const uint8* src_y,
                         uint8* dst_argb,
                         int width) {
   asm volatile (
-    MEMACCESS(4)
-    "vld1.8     {d24}, [%4]                    \n"
-    MEMACCESS(5)
-    "vld1.8     {d25}, [%5]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
+    YUV422TORGB_SETUP_REG
     ".p2align   2                              \n"
   "1:                                          \n"
     READNV12
@@ -615,8 +599,10 @@ void NV12ToARGBRow_NEON(const uint8* src_y,
       "+r"(src_uv),    // %1
       "+r"(dst_argb),  // %2
       "+r"(width)      // %3
-    : "r"(&kUVToRB),   // %4
-      "r"(&kUVToG)     // %5
+    : [kUVToRB]"r"(&kUVToRB),   // %4
+      [kUVToG]"r"(&kUVToG),     // %5
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
     : "cc", "memory", "q0", "q1", "q2", "q3",
       "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
   );
@@ -627,13 +613,7 @@ void NV21ToARGBRow_NEON(const uint8* src_y,
                         uint8* dst_argb,
                         int width) {
   asm volatile (
-    MEMACCESS(4)
-    "vld1.8     {d24}, [%4]                    \n"
-    MEMACCESS(5)
-    "vld1.8     {d25}, [%5]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
+    YUV422TORGB_SETUP_REG
     ".p2align   2                              \n"
   "1:                                          \n"
     READNV21
@@ -647,8 +627,10 @@ void NV21ToARGBRow_NEON(const uint8* src_y,
       "+r"(src_uv),    // %1
       "+r"(dst_argb),  // %2
       "+r"(width)      // %3
-    : "r"(&kUVToRB),   // %4
-      "r"(&kUVToG)     // %5
+    : [kUVToRB]"r"(&kUVToRB),   // %4
+      [kUVToG]"r"(&kUVToG),     // %5
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
     : "cc", "memory", "q0", "q1", "q2", "q3",
       "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
   );
@@ -659,13 +641,7 @@ void NV12ToRGB565Row_NEON(const uint8* src_y,
                           uint8* dst_rgb565,
                           int width) {
   asm volatile (
-    MEMACCESS(4)
-    "vld1.8     {d24}, [%4]                    \n"
-    MEMACCESS(5)
-    "vld1.8     {d25}, [%5]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
+    YUV422TORGB_SETUP_REG
     ".p2align   2                              \n"
   "1:                                          \n"
     READNV12
@@ -679,8 +655,10 @@ void NV12ToRGB565Row_NEON(const uint8* src_y,
       "+r"(src_uv),    // %1
       "+r"(dst_rgb565),  // %2
       "+r"(width)      // %3
-    : "r"(&kUVToRB),   // %4
-      "r"(&kUVToG)     // %5
+    : [kUVToRB]"r"(&kUVToRB),   // %4
+      [kUVToG]"r"(&kUVToG),     // %5
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
     : "cc", "memory", "q0", "q1", "q2", "q3",
       "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
   );
@@ -691,13 +669,7 @@ void NV21ToRGB565Row_NEON(const uint8* src_y,
                           uint8* dst_rgb565,
                           int width) {
   asm volatile (
-    MEMACCESS(4)
-    "vld1.8     {d24}, [%4]                    \n"
-    MEMACCESS(5)
-    "vld1.8     {d25}, [%5]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
+    YUV422TORGB_SETUP_REG
     ".p2align   2                              \n"
   "1:                                          \n"
     READNV21
@@ -711,8 +683,10 @@ void NV21ToRGB565Row_NEON(const uint8* src_y,
       "+r"(src_uv),    // %1
       "+r"(dst_rgb565),  // %2
       "+r"(width)      // %3
-    : "r"(&kUVToRB),   // %4
-      "r"(&kUVToG)     // %5
+    : [kUVToRB]"r"(&kUVToRB),   // %4
+      [kUVToG]"r"(&kUVToG),     // %5
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
     : "cc", "memory", "q0", "q1", "q2", "q3",
       "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
   );
@@ -722,13 +696,7 @@ void YUY2ToARGBRow_NEON(const uint8* src_yuy2,
                         uint8* dst_argb,
                         int width) {
   asm volatile (
-    MEMACCESS(3)
-    "vld1.8     {d24}, [%3]                    \n"
-    MEMACCESS(4)
-    "vld1.8     {d25}, [%4]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
+    YUV422TORGB_SETUP_REG
     ".p2align   2                              \n"
   "1:                                          \n"
     READYUY2
@@ -741,8 +709,10 @@ void YUY2ToARGBRow_NEON(const uint8* src_yuy2,
     : "+r"(src_yuy2),  // %0
       "+r"(dst_argb),  // %1
       "+r"(width)      // %2
-    : "r"(&kUVToRB),   // %3
-      "r"(&kUVToG)     // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %3
+      [kUVToG]"r"(&kUVToG),     // %4
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
     : "cc", "memory", "q0", "q1", "q2", "q3",
       "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
   );
@@ -752,13 +722,7 @@ void UYVYToARGBRow_NEON(const uint8* src_uyvy,
                         uint8* dst_argb,
                         int width) {
   asm volatile (
-    MEMACCESS(3)
-    "vld1.8     {d24}, [%3]                    \n"
-    MEMACCESS(4)
-    "vld1.8     {d25}, [%4]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
+    YUV422TORGB_SETUP_REG
     ".p2align   2                              \n"
   "1:                                          \n"
     READUYVY
@@ -771,8 +735,10 @@ void UYVYToARGBRow_NEON(const uint8* src_uyvy,
     : "+r"(src_uyvy),  // %0
       "+r"(dst_argb),  // %1
       "+r"(width)      // %2
-    : "r"(&kUVToRB),   // %3
-      "r"(&kUVToG)     // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %3
+      [kUVToG]"r"(&kUVToG),     // %4
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
     : "cc", "memory", "q0", "q1", "q2", "q3",
       "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
   );
@@ -844,30 +810,36 @@ void CopyRow_NEON(const uint8* src, uint8* dst, int count) {
   );
 }
 
-// SetRow8 writes 'count' bytes using a 32 bit value repeated.
-void SetRow_NEON(uint8* dst, uint32 v32, int count) {
+// SetRow writes 'count' bytes using an 8 bit value repeated.
+void SetRow_NEON(uint8* dst, uint8 v8, int count) {
   asm volatile (
-    "vdup.u32  q0, %2                          \n"  // duplicate 4 ints
-    "1:                                        \n"
+    "vdup.8    q0, %2                          \n"  // duplicate 16 bytes
+  "1:                                          \n"
     "subs      %1, %1, #16                     \n"  // 16 bytes per loop
     MEMACCESS(0)
     "vst1.8    {q0}, [%0]!                     \n"  // store
     "bgt       1b                              \n"
   : "+r"(dst),   // %0
     "+r"(count)  // %1
-  : "r"(v32)     // %2
+  : "r"(v8)      // %2
   : "cc", "memory", "q0"
   );
 }
 
-// TODO(fbarchard): Make fully assembler
-// SetRow32 writes 'count' words using a 32 bit value repeated.
-void ARGBSetRows_NEON(uint8* dst, uint32 v32, int width,
-                      int dst_stride, int height) {
-  for (int y = 0; y < height; ++y) {
-    SetRow_NEON(dst, v32, width << 2);
-    dst += dst_stride;
-  }
+// ARGBSetRow writes 'count' pixels using an 32 bit value repeated.
+void ARGBSetRow_NEON(uint8* dst, uint32 v32, int count) {
+  asm volatile (
+    "vdup.u32  q0, %2                          \n"  // duplicate 4 ints
+  "1:                                          \n"
+    "subs      %1, %1, #4                      \n"  // 4 pixels per loop
+    MEMACCESS(0)
+    "vst1.8    {q0}, [%0]!                     \n"  // store
+    "bgt       1b                              \n"
+  : "+r"(dst),   // %0
+    "+r"(count)  // %1
+  : "r"(v32)     // %2
+  : "cc", "memory", "q0"
+  );
 }
 
 void MirrorRow_NEON(const uint8* src, uint8* dst, int width) {
@@ -1273,53 +1245,6 @@ void UYVYToUVRow_NEON(const uint8* src_uyvy, int stride_uyvy,
   );
 }
 
-void HalfRow_NEON(const uint8* src_uv, int src_uv_stride,
-                  uint8* dst_uv, int pix) {
-  asm volatile (
-    // change the stride to row 2 pointer
-    "add        %1, %0                         \n"
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // load row 1 16 pixels.
-    "subs       %3, %3, #16                    \n"  // 16 processed per loop
-    MEMACCESS(1)
-    "vld1.8     {q1}, [%1]!                    \n"  // load row 2 16 pixels.
-    "vrhadd.u8  q0, q1                         \n"  // average row 1 and 2
-    MEMACCESS(2)
-    "vst1.8     {q0}, [%2]!                    \n"
-    "bgt        1b                             \n"
-  : "+r"(src_uv),         // %0
-    "+r"(src_uv_stride),  // %1
-    "+r"(dst_uv),         // %2
-    "+r"(pix)             // %3
-  :
-  : "cc", "memory", "q0", "q1"  // Clobber List
-  );
-}
-
-// Select 2 channels from ARGB on alternating pixels.  e.g.  BGBGBGBG
-void ARGBToBayerRow_NEON(const uint8* src_argb, uint8* dst_bayer,
-                         uint32 selector, int pix) {
-  asm volatile (
-    "vmov.u32   d6[0], %3                      \n"  // selector
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {q0, q1}, [%0]!                \n"  // load row 8 pixels.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop
-    "vtbl.8     d4, {d0, d1}, d6               \n"  // look up 4 pixels
-    "vtbl.8     d5, {d2, d3}, d6               \n"  // look up 4 pixels
-    "vtrn.u32   d4, d5                         \n"  // combine 8 pixels
-    MEMACCESS(1)
-    "vst1.8     {d4}, [%1]!                    \n"  // store 8.
-    "bgt        1b                             \n"
-  : "+r"(src_argb),   // %0
-    "+r"(dst_bayer),  // %1
-    "+r"(pix)         // %2
-  : "r"(selector)     // %3
-  : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
-  );
-}
-
 // Select G channels from ARGB.  e.g.  GGGGGGGG
 void ARGBToBayerGGRow_NEON(const uint8* src_argb, uint8* dst_bayer,
                            uint32 /*selector*/, int pix) {
@@ -2832,7 +2757,7 @@ void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb,
     "vmovl.u8   q8, d16                        \n"  // b (0 .. 255) 16 bit
     "vmovl.u8   q9, d18                        \n"  // g
     "vmovl.u8   q10, d20                       \n"  // r
-    "vmovl.u8   q15, d22                       \n"  // a
+    "vmovl.u8   q11, d22                       \n"  // a
     "vmul.s16   q12, q8, d0[0]                 \n"  // B = B * Matrix B
     "vmul.s16   q13, q8, d1[0]                 \n"  // G = B * Matrix G
     "vmul.s16   q14, q8, d2[0]                 \n"  // R = B * Matrix R
@@ -2853,10 +2778,10 @@ void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb,
     "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
     "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
     "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
-    "vmul.s16   q4, q15, d0[3]                 \n"  // B += A * Matrix B
-    "vmul.s16   q5, q15, d1[3]                 \n"  // G += A * Matrix G
-    "vmul.s16   q6, q15, d2[3]                 \n"  // R += A * Matrix R
-    "vmul.s16   q7, q15, d3[3]                 \n"  // A += A * Matrix A
+    "vmul.s16   q4, q11, d0[3]                 \n"  // B += A * Matrix B
+    "vmul.s16   q5, q11, d1[3]                 \n"  // G += A * Matrix G
+    "vmul.s16   q6, q11, d2[3]                 \n"  // R += A * Matrix R
+    "vmul.s16   q7, q11, d3[3]                 \n"  // A += A * Matrix A
     "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
     "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
     "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
@@ -2872,7 +2797,7 @@ void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb,
     "+r"(dst_argb),   // %1
     "+r"(width)       // %2
   : "r"(matrix_argb)  // %3
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9",
+  : "cc", "memory", "q0", "q1", "q2", "q4", "q5", "q6", "q7", "q8", "q9",
     "q10", "q11", "q12", "q13", "q14", "q15"
   );
 }
@@ -3140,7 +3065,7 @@ void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1,
   : "cc", "memory", "q0", "q1"  // Clobber List
   );
 }
-#endif  // __ARM_NEON__
+#endif  // defined(__ARM_NEON__) && !defined(__aarch64__)
 
 #ifdef __cplusplus
 }  // extern "C"
diff --git a/third_party/libyuv/source/row_neon64.cc b/third_party/libyuv/source/row_neon64.cc
index 21111cf60..ddccd5d98 100644
--- a/third_party/libyuv/source/row_neon64.cc
+++ b/third_party/libyuv/source/row_neon64.cc
@@ -1,5 +1,5 @@
 /*
- *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *  Copyright 2014 The LibYuv Project Authors. All rights reserved.
  *
  *  Use of this source code is governed by a BSD-style license
  *  that can be found in the LICENSE file in the root of the source
@@ -15,113 +15,157 @@ namespace libyuv {
 extern "C" {
 #endif
 
-// This module is for GCC Neon
+// This module is for GCC Neon armv8 64 bit.
 #if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
 
 // Read 8 Y, 4 U and 4 V from 422
 #define READYUV422                                                             \
     MEMACCESS(0)                                                               \
-    "vld1.8     {d0}, [%0]!                    \n"                             \
+    "ld1        {v0.8b}, [%0], #8              \n"                             \
     MEMACCESS(1)                                                               \
-    "vld1.32    {d2[0]}, [%1]!                 \n"                             \
+    "ld1        {v1.s}[0], [%1], #4            \n"                             \
     MEMACCESS(2)                                                               \
-    "vld1.32    {d2[1]}, [%2]!                 \n"
+    "ld1        {v1.s}[1], [%2], #4            \n"
 
 // Read 8 Y, 2 U and 2 V from 422
 #define READYUV411                                                             \
     MEMACCESS(0)                                                               \
-    "vld1.8     {d0}, [%0]!                    \n"                             \
+    "ld1        {v0.8b}, [%0], #8              \n"                             \
     MEMACCESS(1)                                                               \
-    "vld1.16    {d2[0]}, [%1]!                 \n"                             \
+    "ld1        {v2.h}[0], [%1], #2            \n"                             \
     MEMACCESS(2)                                                               \
-    "vld1.16    {d2[1]}, [%2]!                 \n"                             \
-    "vmov.u8    d3, d2                         \n"                             \
-    "vzip.u8    d2, d3                         \n"
+    "ld1        {v2.h}[1], [%2], #2            \n"                             \
+    "zip1       v1.8b, v2.8b, v2.8b            \n"
 
 // Read 8 Y, 8 U and 8 V from 444
 #define READYUV444                                                             \
     MEMACCESS(0)                                                               \
-    "vld1.8     {d0}, [%0]!                    \n"                             \
+    "ld1        {v0.8b}, [%0], #8              \n"                             \
     MEMACCESS(1)                                                               \
-    "vld1.8     {d2}, [%1]!                    \n"                             \
+    "ld1        {v1.d}[0], [%1], #8            \n"                             \
     MEMACCESS(2)                                                               \
-    "vld1.8     {d3}, [%2]!                    \n"                             \
-    "vpaddl.u8  q1, q1                         \n"                             \
-    "vrshrn.u16 d2, q1, #1                     \n"
+    "ld1        {v1.d}[1], [%2], #8            \n"                             \
+    "uaddlp     v1.8h, v1.16b                  \n"                             \
+    "rshrn      v1.8b, v1.8h, #1               \n"
 
 // Read 8 Y, and set 4 U and 4 V to 128
 #define READYUV400                                                             \
     MEMACCESS(0)                                                               \
-    "vld1.8     {d0}, [%0]!                    \n"                             \
-    "vmov.u8    d2, #128                       \n"
+    "ld1        {v0.8b}, [%0], #8              \n"                             \
+    "movi       v1.8b , #128                   \n"
 
 // Read 8 Y and 4 UV from NV12
 #define READNV12                                                               \
     MEMACCESS(0)                                                               \
-    "vld1.8     {d0}, [%0]!                    \n"                             \
+    "ld1        {v0.8b}, [%0], #8              \n"                             \
     MEMACCESS(1)                                                               \
-    "vld1.8     {d2}, [%1]!                    \n"                             \
-    "vmov.u8    d3, d2                         \n"/* split odd/even uv apart */\
-    "vuzp.u8    d2, d3                         \n"                             \
-    "vtrn.u32   d2, d3                         \n"
+    "ld1        {v2.8b}, [%1], #8              \n"                             \
+    "uzp1       v1.8b, v2.8b, v2.8b            \n"                             \
+    "uzp2       v3.8b, v2.8b, v2.8b            \n"                             \
+    "ins        v1.s[1], v3.s[0]               \n"
 
 // Read 8 Y and 4 VU from NV21
 #define READNV21                                                               \
     MEMACCESS(0)                                                               \
-    "vld1.8     {d0}, [%0]!                    \n"                             \
+    "ld1        {v0.8b}, [%0], #8              \n"                             \
     MEMACCESS(1)                                                               \
-    "vld1.8     {d2}, [%1]!                    \n"                             \
-    "vmov.u8    d3, d2                         \n"/* split odd/even uv apart */\
-    "vuzp.u8    d3, d2                         \n"                             \
-    "vtrn.u32   d2, d3                         \n"
+    "ld1        {v2.8b}, [%1], #8              \n"                             \
+    "uzp1       v3.8b, v2.8b, v2.8b            \n"                             \
+    "uzp2       v1.8b, v2.8b, v2.8b            \n"                             \
+    "ins        v1.s[1], v3.s[0]               \n"
 
 // Read 8 YUY2
 #define READYUY2                                                               \
     MEMACCESS(0)                                                               \
-    "vld2.8     {d0, d2}, [%0]!                \n"                             \
-    "vmov.u8    d3, d2                         \n"                             \
-    "vuzp.u8    d2, d3                         \n"                             \
-    "vtrn.u32   d2, d3                         \n"
+    "ld2        {v0.8b, v1.8b}, [%0], #16      \n"                             \
+    "uzp2       v3.8b, v1.8b, v1.8b            \n"                             \
+    "uzp1       v1.8b, v1.8b, v1.8b            \n"                             \
+    "ins        v1.s[1], v3.s[0]               \n"
 
 // Read 8 UYVY
 #define READUYVY                                                               \
     MEMACCESS(0)                                                               \
-    "vld2.8     {d2, d3}, [%0]!                \n"                             \
-    "vmov.u8    d0, d3                         \n"                             \
-    "vmov.u8    d3, d2                         \n"                             \
-    "vuzp.u8    d2, d3                         \n"                             \
-    "vtrn.u32   d2, d3                         \n"
-
-#define YUV422TORGB                                                            \
-    "veor.u8    d2, d26                        \n"/*subtract 128 from u and v*/\
-    "vmull.s8   q8, d2, d24                    \n"/*  u/v B/R component      */\
-    "vmull.s8   q9, d2, d25                    \n"/*  u/v G component        */\
-    "vmov.u8    d1, #0                         \n"/*  split odd/even y apart */\
-    "vtrn.u8    d0, d1                         \n"                             \
-    "vsub.s16   q0, q0, q15                    \n"/*  offset y               */\
-    "vmul.s16   q0, q0, q14                    \n"                             \
-    "vadd.s16   d18, d19                       \n"                             \
-    "vqadd.s16  d20, d0, d16                   \n" /* B */                     \
-    "vqadd.s16  d21, d1, d16                   \n"                             \
-    "vqadd.s16  d22, d0, d17                   \n" /* R */                     \
-    "vqadd.s16  d23, d1, d17                   \n"                             \
-    "vqadd.s16  d16, d0, d18                   \n" /* G */                     \
-    "vqadd.s16  d17, d1, d18                   \n"                             \
-    "vqshrun.s16 d0, q10, #6                   \n" /* B */                     \
-    "vqshrun.s16 d1, q11, #6                   \n" /* G */                     \
-    "vqshrun.s16 d2, q8, #6                    \n" /* R */                     \
-    "vmovl.u8   q10, d0                        \n"/*  set up for reinterleave*/\
-    "vmovl.u8   q11, d1                        \n"                             \
-    "vmovl.u8   q8, d2                         \n"                             \
-    "vtrn.u8    d20, d21                       \n"                             \
-    "vtrn.u8    d22, d23                       \n"                             \
-    "vtrn.u8    d16, d17                       \n"                             \
-    "vmov.u8    d21, d16                       \n"
-
-static vec8 kUVToRB  = { 127, 127, 127, 127, 102, 102, 102, 102,
-                         0, 0, 0, 0, 0, 0, 0, 0 };
-static vec8 kUVToG = { -25, -25, -25, -25, -52, -52, -52, -52,
-                       0, 0, 0, 0, 0, 0, 0, 0 };
+    "ld2        {v2.8b, v3.8b}, [%0], #16      \n"                             \
+    "orr        v0.8b, v3.8b, v3.8b            \n"                             \
+    "uzp1       v1.8b, v2.8b, v2.8b            \n"                             \
+    "uzp2       v3.8b, v2.8b, v2.8b            \n"                             \
+    "ins        v1.s[1], v3.s[0]               \n"
+
+#define YUV422TORGB_SETUP_REG                                                  \
+    "ld1r       {v24.8h}, [%[kUVBiasBGR]], #2  \n"                             \
+    "ld1r       {v25.8h}, [%[kUVBiasBGR]], #2  \n"                             \
+    "ld1r       {v26.8h}, [%[kUVBiasBGR]]      \n"                             \
+    "ld1r       {v31.4s}, [%[kYToRgb]]         \n"                             \
+    "movi       v27.8h, #128                   \n"                             \
+    "movi       v28.8h, #102                   \n"                             \
+    "movi       v29.8h, #25                    \n"                             \
+    "movi       v30.8h, #52                    \n"
+
+#define YUV422TORGB(vR, vG, vB)                                                \
+    "uxtl       v0.8h, v0.8b                   \n" /* Extract Y    */          \
+    "shll       v2.8h, v1.8b, #8               \n" /* Replicate UV */          \
+    "ushll2     v3.4s, v0.8h, #0               \n" /* Y */                     \
+    "ushll      v0.4s, v0.4h, #0               \n"                             \
+    "mul        v3.4s, v3.4s, v31.4s           \n"                             \
+    "mul        v0.4s, v0.4s, v31.4s           \n"                             \
+    "sqshrun    v0.4h, v0.4s, #16              \n"                             \
+    "sqshrun2   v0.8h, v3.4s, #16              \n" /* Y */                     \
+    "uaddw      v1.8h, v2.8h, v1.8b            \n" /* Replicate UV */          \
+    "mov        v2.d[0], v1.d[1]               \n" /* Extract V */             \
+    "uxtl       v2.8h, v2.8b                   \n"                             \
+    "uxtl       v1.8h, v1.8b                   \n" /* Extract U */             \
+    "mul        v3.8h, v1.8h, v27.8h           \n"                             \
+    "mul        v5.8h, v1.8h, v29.8h           \n"                             \
+    "mul        v6.8h, v2.8h, v30.8h           \n"                             \
+    "mul        v7.8h, v2.8h, v28.8h           \n"                             \
+    "sqadd      v6.8h, v6.8h, v5.8h            \n"                             \
+    "sqadd      " #vB ".8h, v24.8h, v0.8h      \n" /* B */                     \
+    "sqadd      " #vG ".8h, v25.8h, v0.8h      \n" /* G */                     \
+    "sqadd      " #vR ".8h, v26.8h, v0.8h      \n" /* R */                     \
+    "sqadd      " #vB ".8h, " #vB ".8h, v3.8h  \n" /* B */                     \
+    "sqsub      " #vG ".8h, " #vG ".8h, v6.8h  \n" /* G */                     \
+    "sqadd      " #vR ".8h, " #vR ".8h, v7.8h  \n" /* R */                     \
+    "sqshrun    " #vB ".8b, " #vB ".8h, #6     \n" /* B */                     \
+    "sqshrun    " #vG ".8b, " #vG ".8h, #6     \n" /* G */                     \
+    "sqshrun    " #vR ".8b, " #vR ".8h, #6     \n" /* R */                     \
+
+// YUV to RGB conversion constants.
+// Y contribution to R,G,B.  Scale and bias.
+#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */
+#define YGB 1160 /* 1.164 * 64 * 16 - adjusted for even error distribution */
+
+// U and V contributions to R,G,B.
+#define UB -128 /* -min(128, round(2.018 * 64)) */
+#define UG 25 /* -round(-0.391 * 64) */
+#define VG 52 /* -round(-0.813 * 64) */
+#define VR -102 /* -round(1.596 * 64) */
+
+// Bias values to subtract 16 from Y and 128 from U and V.
+#define BB (UB * 128            - YGB)
+#define BG (UG * 128 + VG * 128 - YGB)
+#define BR            (VR * 128 - YGB)
+
+static vec16 kUVBiasBGR = { BB, BG, BR, 0, 0, 0, 0, 0 };
+static vec32 kYToRgb = { 0x0101 * YG, 0, 0, 0 };
+
+#undef YG
+#undef YGB
+#undef UB
+#undef UG
+#undef VG
+#undef VR
+#undef BB
+#undef BG
+#undef BR
+
+#define RGBTOUV_SETUP_REG                                                      \
+    "movi       v20.8h, #56, lsl #0  \n"  /* UB/VR coefficient (0.875) / 2 */  \
+    "movi       v21.8h, #37, lsl #0  \n"  /* UG coefficient (-0.5781) / 2  */  \
+    "movi       v22.8h, #19, lsl #0  \n"  /* UR coefficient (-0.2969) / 2  */  \
+    "movi       v23.8h, #9,  lsl #0  \n"  /* VB coefficient (-0.1406) / 2  */  \
+    "movi       v24.8h, #47, lsl #0  \n"  /* VG coefficient (-0.7344) / 2  */  \
+    "movi       v25.16b, #0x80       \n"  /* 128.5 (0x8080 in 16-bit)      */
+
 
 #ifdef HAS_I444TOARGBROW_NEON
 void I444ToARGBRow_NEON(const uint8* src_y,
@@ -130,31 +174,24 @@ void I444ToARGBRow_NEON(const uint8* src_y,
                         uint8* dst_argb,
                         int width) {
   asm volatile (
-    MEMACCESS(5)
-    "vld1.8     {d24}, [%5]                    \n"
-    MEMACCESS(6)
-    "vld1.8     {d25}, [%6]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
-    ".p2align   2                              \n"
+    YUV422TORGB_SETUP_REG
   "1:                                          \n"
     READYUV444
-    YUV422TORGB
+    YUV422TORGB(v22, v21, v20)
     "subs       %4, %4, #8                     \n"
-    "vmov.u8    d23, #255                      \n"
+    "movi       v23.8b, #255                   \n" /* A */
     MEMACCESS(3)
-    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
-    "bgt        1b                             \n"
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32 \n"
+    "b.gt       1b                             \n"
     : "+r"(src_y),     // %0
       "+r"(src_u),     // %1
       "+r"(src_v),     // %2
       "+r"(dst_argb),  // %3
       "+r"(width)      // %4
-    : "r"(&kUVToRB),   // %5
-      "r"(&kUVToG)     // %6
-    : "cc", "memory", "q0", "q1", "q2", "q3",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
   );
 }
 #endif  // HAS_I444TOARGBROW_NEON
@@ -166,31 +203,24 @@ void I422ToARGBRow_NEON(const uint8* src_y,
                         uint8* dst_argb,
                         int width) {
   asm volatile (
-    MEMACCESS(5)
-    "vld1.8     {d24}, [%5]                    \n"
-    MEMACCESS(6)
-    "vld1.8     {d25}, [%6]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
-    ".p2align   2                              \n"
+    YUV422TORGB_SETUP_REG
   "1:                                          \n"
     READYUV422
-    YUV422TORGB
+    YUV422TORGB(v22, v21, v20)
     "subs       %4, %4, #8                     \n"
-    "vmov.u8    d23, #255                      \n"
+    "movi       v23.8b, #255                   \n" /* A */
     MEMACCESS(3)
-    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
-    "bgt        1b                             \n"
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32     \n"
+    "b.gt       1b                             \n"
     : "+r"(src_y),     // %0
       "+r"(src_u),     // %1
       "+r"(src_v),     // %2
       "+r"(dst_argb),  // %3
       "+r"(width)      // %4
-    : "r"(&kUVToRB),   // %5
-      "r"(&kUVToG)     // %6
-    : "cc", "memory", "q0", "q1", "q2", "q3",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
   );
 }
 #endif  // HAS_I422TOARGBROW_NEON
@@ -202,31 +232,24 @@ void I411ToARGBRow_NEON(const uint8* src_y,
                         uint8* dst_argb,
                         int width) {
   asm volatile (
-    MEMACCESS(5)
-    "vld1.8     {d24}, [%5]                    \n"
-    MEMACCESS(6)
-    "vld1.8     {d25}, [%6]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
-    ".p2align   2                              \n"
+    YUV422TORGB_SETUP_REG
   "1:                                          \n"
     READYUV411
-    YUV422TORGB
+    YUV422TORGB(v22, v21, v20)
     "subs       %4, %4, #8                     \n"
-    "vmov.u8    d23, #255                      \n"
+    "movi       v23.8b, #255                   \n" /* A */
     MEMACCESS(3)
-    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
-    "bgt        1b                             \n"
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32     \n"
+    "b.gt       1b                             \n"
     : "+r"(src_y),     // %0
       "+r"(src_u),     // %1
       "+r"(src_v),     // %2
       "+r"(dst_argb),  // %3
       "+r"(width)      // %4
-    : "r"(&kUVToRB),   // %5
-      "r"(&kUVToG)     // %6
-    : "cc", "memory", "q0", "q1", "q2", "q3",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
   );
 }
 #endif  // HAS_I411TOARGBROW_NEON
@@ -238,32 +261,24 @@ void I422ToBGRARow_NEON(const uint8* src_y,
                         uint8* dst_bgra,
                         int width) {
   asm volatile (
-    MEMACCESS(5)
-    "vld1.8     {d24}, [%5]                    \n"
-    MEMACCESS(6)
-    "vld1.8     {d25}, [%6]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
-    ".p2align   2                              \n"
+    YUV422TORGB_SETUP_REG
   "1:                                          \n"
     READYUV422
-    YUV422TORGB
+    YUV422TORGB(v21, v22, v23)
     "subs       %4, %4, #8                     \n"
-    "vswp.u8    d20, d22                       \n"
-    "vmov.u8    d19, #255                      \n"
+    "movi       v20.8b, #255                   \n" /* A */
     MEMACCESS(3)
-    "vst4.8     {d19, d20, d21, d22}, [%3]!    \n"
-    "bgt        1b                             \n"
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32     \n"
+    "b.gt       1b                             \n"
     : "+r"(src_y),     // %0
       "+r"(src_u),     // %1
       "+r"(src_v),     // %2
       "+r"(dst_bgra),  // %3
       "+r"(width)      // %4
-    : "r"(&kUVToRB),   // %5
-      "r"(&kUVToG)     // %6
-    : "cc", "memory", "q0", "q1", "q2", "q3",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
   );
 }
 #endif  // HAS_I422TOBGRAROW_NEON
@@ -275,32 +290,24 @@ void I422ToABGRRow_NEON(const uint8* src_y,
                         uint8* dst_abgr,
                         int width) {
   asm volatile (
-    MEMACCESS(5)
-    "vld1.8     {d24}, [%5]                    \n"
-    MEMACCESS(6)
-    "vld1.8     {d25}, [%6]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
-    ".p2align   2                              \n"
+    YUV422TORGB_SETUP_REG
   "1:                                          \n"
     READYUV422
-    YUV422TORGB
+    YUV422TORGB(v20, v21, v22)
     "subs       %4, %4, #8                     \n"
-    "vswp.u8    d20, d22                       \n"
-    "vmov.u8    d23, #255                      \n"
+    "movi       v23.8b, #255                   \n" /* A */
     MEMACCESS(3)
-    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
-    "bgt        1b                             \n"
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32     \n"
+    "b.gt       1b                             \n"
     : "+r"(src_y),     // %0
       "+r"(src_u),     // %1
       "+r"(src_v),     // %2
       "+r"(dst_abgr),  // %3
       "+r"(width)      // %4
-    : "r"(&kUVToRB),   // %5
-      "r"(&kUVToG)     // %6
-    : "cc", "memory", "q0", "q1", "q2", "q3",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
   );
 }
 #endif  // HAS_I422TOABGRROW_NEON
@@ -312,31 +319,24 @@ void I422ToRGBARow_NEON(const uint8* src_y,
                         uint8* dst_rgba,
                         int width) {
   asm volatile (
-    MEMACCESS(5)
-    "vld1.8     {d24}, [%5]                    \n"
-    MEMACCESS(6)
-    "vld1.8     {d25}, [%6]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
-    ".p2align   2                              \n"
+    YUV422TORGB_SETUP_REG
   "1:                                          \n"
     READYUV422
-    YUV422TORGB
+    YUV422TORGB(v23, v22, v21)
     "subs       %4, %4, #8                     \n"
-    "vmov.u8    d19, #255                      \n"
+    "movi       v20.8b, #255                   \n" /* A */
     MEMACCESS(3)
-    "vst4.8     {d19, d20, d21, d22}, [%3]!    \n"
-    "bgt        1b                             \n"
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32     \n"
+    "b.gt       1b                             \n"
     : "+r"(src_y),     // %0
       "+r"(src_u),     // %1
       "+r"(src_v),     // %2
       "+r"(dst_rgba),  // %3
       "+r"(width)      // %4
-    : "r"(&kUVToRB),   // %5
-      "r"(&kUVToG)     // %6
-    : "cc", "memory", "q0", "q1", "q2", "q3",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
   );
 }
 #endif  // HAS_I422TORGBAROW_NEON
@@ -348,30 +348,23 @@ void I422ToRGB24Row_NEON(const uint8* src_y,
                          uint8* dst_rgb24,
                          int width) {
   asm volatile (
-    MEMACCESS(5)
-    "vld1.8     {d24}, [%5]                    \n"
-    MEMACCESS(6)
-    "vld1.8     {d25}, [%6]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
-    ".p2align   2                              \n"
+    YUV422TORGB_SETUP_REG
   "1:                                          \n"
     READYUV422
-    YUV422TORGB
+    YUV422TORGB(v22, v21, v20)
     "subs       %4, %4, #8                     \n"
     MEMACCESS(3)
-    "vst3.8     {d20, d21, d22}, [%3]!         \n"
-    "bgt        1b                             \n"
-    : "+r"(src_y),      // %0
-      "+r"(src_u),      // %1
-      "+r"(src_v),      // %2
-      "+r"(dst_rgb24),  // %3
-      "+r"(width)       // %4
-    : "r"(&kUVToRB),    // %5
-      "r"(&kUVToG)      // %6
-    : "cc", "memory", "q0", "q1", "q2", "q3",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+    "st3        {v20.8b,v21.8b,v22.8b}, [%3], #24     \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_rgb24), // %3
+      "+r"(width)      // %4
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
   );
 }
 #endif  // HAS_I422TORGB24ROW_NEON
@@ -383,46 +376,33 @@ void I422ToRAWRow_NEON(const uint8* src_y,
                        uint8* dst_raw,
                        int width) {
   asm volatile (
-    MEMACCESS(5)
-    "vld1.8     {d24}, [%5]                    \n"
-    MEMACCESS(6)
-    "vld1.8     {d25}, [%6]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
-    ".p2align   2                              \n"
+    YUV422TORGB_SETUP_REG
   "1:                                          \n"
     READYUV422
-    YUV422TORGB
+    YUV422TORGB(v20, v21, v22)
     "subs       %4, %4, #8                     \n"
-    "vswp.u8    d20, d22                       \n"
     MEMACCESS(3)
-    "vst3.8     {d20, d21, d22}, [%3]!         \n"
-    "bgt        1b                             \n"
-    : "+r"(src_y),    // %0
-      "+r"(src_u),    // %1
-      "+r"(src_v),    // %2
-      "+r"(dst_raw),  // %3
-      "+r"(width)     // %4
-    : "r"(&kUVToRB),  // %5
-      "r"(&kUVToG)    // %6
-    : "cc", "memory", "q0", "q1", "q2", "q3",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+    "st3        {v20.8b,v21.8b,v22.8b}, [%3], #24     \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_raw),   // %3
+      "+r"(width)      // %4
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
   );
 }
 #endif  // HAS_I422TORAWROW_NEON
 
 #define ARGBTORGB565                                                           \
-    "vshr.u8    d20, d20, #3                   \n"  /* B                    */ \
-    "vshr.u8    d21, d21, #2                   \n"  /* G                    */ \
-    "vshr.u8    d22, d22, #3                   \n"  /* R                    */ \
-    "vmovl.u8   q8, d20                        \n"  /* B                    */ \
-    "vmovl.u8   q9, d21                        \n"  /* G                    */ \
-    "vmovl.u8   q10, d22                       \n"  /* R                    */ \
-    "vshl.u16   q9, q9, #5                     \n"  /* G                    */ \
-    "vshl.u16   q10, q10, #11                  \n"  /* R                    */ \
-    "vorr       q0, q8, q9                     \n"  /* BG                   */ \
-    "vorr       q0, q0, q10                    \n"  /* BGR                  */
+    "shll       v0.8h,  v22.8b, #8             \n"  /* R                    */ \
+    "shll       v20.8h, v20.8b, #8             \n"  /* B                    */ \
+    "shll       v21.8h, v21.8b, #8             \n"  /* G                    */ \
+    "sri        v0.8h,  v21.8h, #5             \n"  /* RG                   */ \
+    "sri        v0.8h,  v20.8h, #11            \n"  /* RGB                  */
 
 #ifdef HAS_I422TORGB565ROW_NEON
 void I422ToRGB565Row_NEON(const uint8* src_y,
@@ -431,49 +411,36 @@ void I422ToRGB565Row_NEON(const uint8* src_y,
                           uint8* dst_rgb565,
                           int width) {
   asm volatile (
-    MEMACCESS(5)
-    "vld1.8     {d24}, [%5]                    \n"
-    MEMACCESS(6)
-    "vld1.8     {d25}, [%6]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
-    ".p2align   2                              \n"
+    YUV422TORGB_SETUP_REG
   "1:                                          \n"
     READYUV422
-    YUV422TORGB
+    YUV422TORGB(v22, v21, v20)
     "subs       %4, %4, #8                     \n"
     ARGBTORGB565
     MEMACCESS(3)
-    "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels RGB565.
-    "bgt        1b                             \n"
+    "st1        {v0.8h}, [%3], #16             \n"  // store 8 pixels RGB565.
+    "b.gt       1b                             \n"
     : "+r"(src_y),    // %0
       "+r"(src_u),    // %1
       "+r"(src_v),    // %2
       "+r"(dst_rgb565),  // %3
       "+r"(width)     // %4
-    : "r"(&kUVToRB),  // %5
-      "r"(&kUVToG)    // %6
-    : "cc", "memory", "q0", "q1", "q2", "q3",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
   );
 }
 #endif  // HAS_I422TORGB565ROW_NEON
 
 #define ARGBTOARGB1555                                                         \
-    "vshr.u8    q10, q10, #3                   \n"  /* B                    */ \
-    "vshr.u8    d22, d22, #3                   \n"  /* R                    */ \
-    "vshr.u8    d23, d23, #7                   \n"  /* A                    */ \
-    "vmovl.u8   q8, d20                        \n"  /* B                    */ \
-    "vmovl.u8   q9, d21                        \n"  /* G                    */ \
-    "vmovl.u8   q10, d22                       \n"  /* R                    */ \
-    "vmovl.u8   q11, d23                       \n"  /* A                    */ \
-    "vshl.u16   q9, q9, #5                     \n"  /* G                    */ \
-    "vshl.u16   q10, q10, #10                  \n"  /* R                    */ \
-    "vshl.u16   q11, q11, #15                  \n"  /* A                    */ \
-    "vorr       q0, q8, q9                     \n"  /* BG                   */ \
-    "vorr       q1, q10, q11                   \n"  /* RA                   */ \
-    "vorr       q0, q0, q1                     \n"  /* BGRA                 */
+    "shll       v0.8h,  v23.8b, #8             \n"  /* A                    */ \
+    "shll       v22.8h, v22.8b, #8             \n"  /* R                    */ \
+    "shll       v20.8h, v20.8b, #8             \n"  /* B                    */ \
+    "shll       v21.8h, v21.8b, #8             \n"  /* G                    */ \
+    "sri        v0.8h,  v22.8h, #1             \n"  /* AR                   */ \
+    "sri        v0.8h,  v21.8h, #6             \n"  /* ARG                  */ \
+    "sri        v0.8h,  v20.8h, #11            \n"  /* ARGB                 */
 
 #ifdef HAS_I422TOARGB1555ROW_NEON
 void I422ToARGB1555Row_NEON(const uint8* src_y,
@@ -482,44 +449,38 @@ void I422ToARGB1555Row_NEON(const uint8* src_y,
                             uint8* dst_argb1555,
                             int width) {
   asm volatile (
-    MEMACCESS(5)
-    "vld1.8     {d24}, [%5]                    \n"
-    MEMACCESS(6)
-    "vld1.8     {d25}, [%6]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
-    ".p2align   2                              \n"
+    YUV422TORGB_SETUP_REG
   "1:                                          \n"
     READYUV422
-    YUV422TORGB
+    YUV422TORGB(v22, v21, v20)
     "subs       %4, %4, #8                     \n"
-    "vmov.u8    d23, #255                      \n"
+    "movi       v23.8b, #255                   \n"
     ARGBTOARGB1555
     MEMACCESS(3)
-    "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels ARGB1555.
-    "bgt        1b                             \n"
+    "st1        {v0.8h}, [%3], #16             \n"  // store 8 pixels RGB565.
+    "b.gt       1b                             \n"
     : "+r"(src_y),    // %0
       "+r"(src_u),    // %1
       "+r"(src_v),    // %2
       "+r"(dst_argb1555),  // %3
       "+r"(width)     // %4
-    : "r"(&kUVToRB),  // %5
-      "r"(&kUVToG)    // %6
-    : "cc", "memory", "q0", "q1", "q2", "q3",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
   );
 }
 #endif  // HAS_I422TOARGB1555ROW_NEON
 
 #define ARGBTOARGB4444                                                         \
-    "vshr.u8    d20, d20, #4                   \n"  /* B                    */ \
-    "vbic.32    d21, d21, d4                   \n"  /* G                    */ \
-    "vshr.u8    d22, d22, #4                   \n"  /* R                    */ \
-    "vbic.32    d23, d23, d4                   \n"  /* A                    */ \
-    "vorr       d0, d20, d21                   \n"  /* BG                   */ \
-    "vorr       d1, d22, d23                   \n"  /* RA                   */ \
-    "vzip.u8    d0, d1                         \n"  /* BGRA                 */
+    /* Input v20.8b<=B, v21.8b<=G, v22.8b<=R, v23.8b<=A, v4.8b<=0x0f        */ \
+    "ushr       v20.8b, v20.8b, #4             \n"  /* B                    */ \
+    "bic        v21.8b, v21.8b, v4.8b          \n"  /* G                    */ \
+    "ushr       v22.8b, v22.8b, #4             \n"  /* R                    */ \
+    "bic        v23.8b, v23.8b, v4.8b          \n"  /* A                    */ \
+    "orr        v0.8b,  v20.8b, v21.8b         \n"  /* BG                   */ \
+    "orr        v1.8b,  v22.8b, v23.8b         \n"  /* RA                   */ \
+    "zip1       v0.16b, v0.16b, v1.16b         \n"  /* BGRA                 */
 
 #ifdef HAS_I422TOARGB4444ROW_NEON
 void I422ToARGB4444Row_NEON(const uint8* src_y,
@@ -528,33 +489,26 @@ void I422ToARGB4444Row_NEON(const uint8* src_y,
                             uint8* dst_argb4444,
                             int width) {
   asm volatile (
-    MEMACCESS(5)
-    "vld1.8     {d24}, [%5]                    \n"
-    MEMACCESS(6)
-    "vld1.8     {d25}, [%6]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
-    "vmov.u8    d4, #0x0f                      \n"  // bits to clear with vbic.
-    ".p2align   2                              \n"
+    YUV422TORGB_SETUP_REG
+    "movi       v4.16b, #0x0f                  \n"  // bits to clear with vbic.
   "1:                                          \n"
     READYUV422
-    YUV422TORGB
+    YUV422TORGB(v22, v21, v20)
     "subs       %4, %4, #8                     \n"
-    "vmov.u8    d23, #255                      \n"
+    "movi       v23.8b, #255                   \n"
     ARGBTOARGB4444
     MEMACCESS(3)
-    "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels ARGB4444.
-    "bgt        1b                             \n"
+    "st1        {v0.8h}, [%3], #16             \n"  // store 8 pixels ARGB4444.
+    "b.gt       1b                             \n"
     : "+r"(src_y),    // %0
       "+r"(src_u),    // %1
       "+r"(src_v),    // %2
       "+r"(dst_argb4444),  // %3
       "+r"(width)     // %4
-    : "r"(&kUVToRB),  // %5
-      "r"(&kUVToG)    // %6
-    : "cc", "memory", "q0", "q1", "q2", "q3",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
   );
 }
 #endif  // HAS_I422TOARGB4444ROW_NEON
@@ -564,29 +518,22 @@ void YToARGBRow_NEON(const uint8* src_y,
                      uint8* dst_argb,
                      int width) {
   asm volatile (
-    MEMACCESS(3)
-    "vld1.8     {d24}, [%3]                    \n"
-    MEMACCESS(4)
-    "vld1.8     {d25}, [%4]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
-    ".p2align   2                              \n"
+    YUV422TORGB_SETUP_REG
   "1:                                          \n"
     READYUV400
-    YUV422TORGB
+    YUV422TORGB(v22, v21, v20)
     "subs       %2, %2, #8                     \n"
-    "vmov.u8    d23, #255                      \n"
+    "movi       v23.8b, #255                   \n"
     MEMACCESS(1)
-    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
-    "bgt        1b                             \n"
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32     \n"
+    "b.gt       1b                             \n"
     : "+r"(src_y),     // %0
       "+r"(dst_argb),  // %1
       "+r"(width)      // %2
-    : "r"(&kUVToRB),   // %3
-      "r"(&kUVToG)     // %4
-    : "cc", "memory", "q0", "q1", "q2", "q3",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
   );
 }
 #endif  // HAS_YTOARGBROW_NEON
@@ -596,22 +543,21 @@ void I400ToARGBRow_NEON(const uint8* src_y,
                         uint8* dst_argb,
                         int width) {
   asm volatile (
-    ".p2align   2                              \n"
-    "vmov.u8    d23, #255                      \n"
+    "movi       v23.8b, #255                   \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "vld1.8     {d20}, [%0]!                   \n"
-    "vmov       d21, d20                       \n"
-    "vmov       d22, d20                       \n"
+    "ld1        {v20.8b}, [%0], #8             \n"
+    "orr        v21.8b, v20.8b, v20.8b         \n"
+    "orr        v22.8b, v20.8b, v20.8b         \n"
     "subs       %2, %2, #8                     \n"
     MEMACCESS(1)
-    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
-    "bgt        1b                             \n"
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32     \n"
+    "b.gt       1b                             \n"
     : "+r"(src_y),     // %0
       "+r"(dst_argb),  // %1
       "+r"(width)      // %2
     :
-    : "cc", "memory", "d20", "d21", "d22", "d23"
+    : "cc", "memory", "v20", "v21", "v22", "v23"
   );
 }
 #endif  // HAS_I400TOARGBROW_NEON
@@ -622,30 +568,23 @@ void NV12ToARGBRow_NEON(const uint8* src_y,
                         uint8* dst_argb,
                         int width) {
   asm volatile (
-    MEMACCESS(4)
-    "vld1.8     {d24}, [%4]                    \n"
-    MEMACCESS(5)
-    "vld1.8     {d25}, [%5]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
-    ".p2align   2                              \n"
+    YUV422TORGB_SETUP_REG
   "1:                                          \n"
     READNV12
-    YUV422TORGB
+    YUV422TORGB(v22, v21, v20)
     "subs       %3, %3, #8                     \n"
-    "vmov.u8    d23, #255                      \n"
+    "movi       v23.8b, #255                   \n"
     MEMACCESS(2)
-    "vst4.8     {d20, d21, d22, d23}, [%2]!    \n"
-    "bgt        1b                             \n"
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%2], #32     \n"
+    "b.gt       1b                             \n"
     : "+r"(src_y),     // %0
       "+r"(src_uv),    // %1
       "+r"(dst_argb),  // %2
       "+r"(width)      // %3
-    : "r"(&kUVToRB),   // %4
-      "r"(&kUVToG)     // %5
-    : "cc", "memory", "q0", "q1", "q2", "q3",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
   );
 }
 #endif  // HAS_NV12TOARGBROW_NEON
@@ -656,30 +595,23 @@ void NV21ToARGBRow_NEON(const uint8* src_y,
                         uint8* dst_argb,
                         int width) {
   asm volatile (
-    MEMACCESS(4)
-    "vld1.8     {d24}, [%4]                    \n"
-    MEMACCESS(5)
-    "vld1.8     {d25}, [%5]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
-    ".p2align   2                              \n"
+    YUV422TORGB_SETUP_REG
   "1:                                          \n"
     READNV21
-    YUV422TORGB
+    YUV422TORGB(v22, v21, v20)
     "subs       %3, %3, #8                     \n"
-    "vmov.u8    d23, #255                      \n"
+    "movi       v23.8b, #255                   \n"
     MEMACCESS(2)
-    "vst4.8     {d20, d21, d22, d23}, [%2]!    \n"
-    "bgt        1b                             \n"
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%2], #32     \n"
+    "b.gt       1b                             \n"
     : "+r"(src_y),     // %0
       "+r"(src_uv),    // %1
       "+r"(dst_argb),  // %2
       "+r"(width)      // %3
-    : "r"(&kUVToRB),   // %4
-      "r"(&kUVToG)     // %5
-    : "cc", "memory", "q0", "q1", "q2", "q3",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
   );
 }
 #endif  // HAS_NV21TOARGBROW_NEON
@@ -690,30 +622,23 @@ void NV12ToRGB565Row_NEON(const uint8* src_y,
                           uint8* dst_rgb565,
                           int width) {
   asm volatile (
-    MEMACCESS(4)
-    "vld1.8     {d24}, [%4]                    \n"
-    MEMACCESS(5)
-    "vld1.8     {d25}, [%5]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
-    ".p2align   2                              \n"
+    YUV422TORGB_SETUP_REG
   "1:                                          \n"
     READNV12
-    YUV422TORGB
+    YUV422TORGB(v22, v21, v20)
     "subs       %3, %3, #8                     \n"
     ARGBTORGB565
     MEMACCESS(2)
-    "vst1.8     {q0}, [%2]!                    \n"  // store 8 pixels RGB565.
-    "bgt        1b                             \n"
+    "st1        {v0.8h}, [%2], 16              \n"  // store 8 pixels RGB565.
+    "b.gt       1b                             \n"
     : "+r"(src_y),     // %0
       "+r"(src_uv),    // %1
       "+r"(dst_rgb565),  // %2
       "+r"(width)      // %3
-    : "r"(&kUVToRB),   // %4
-      "r"(&kUVToG)     // %5
-    : "cc", "memory", "q0", "q1", "q2", "q3",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
   );
 }
 #endif  // HAS_NV12TORGB565ROW_NEON
@@ -724,30 +649,23 @@ void NV21ToRGB565Row_NEON(const uint8* src_y,
                           uint8* dst_rgb565,
                           int width) {
   asm volatile (
-    MEMACCESS(4)
-    "vld1.8     {d24}, [%4]                    \n"
-    MEMACCESS(5)
-    "vld1.8     {d25}, [%5]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
-    ".p2align   2                              \n"
+    YUV422TORGB_SETUP_REG
   "1:                                          \n"
     READNV21
-    YUV422TORGB
+    YUV422TORGB(v22, v21, v20)
     "subs       %3, %3, #8                     \n"
     ARGBTORGB565
     MEMACCESS(2)
-    "vst1.8     {q0}, [%2]!                    \n"  // store 8 pixels RGB565.
-    "bgt        1b                             \n"
+    "st1        {v0.8h}, [%2], 16              \n"  // store 8 pixels RGB565.
+    "b.gt       1b                             \n"
     : "+r"(src_y),     // %0
       "+r"(src_uv),    // %1
       "+r"(dst_rgb565),  // %2
       "+r"(width)      // %3
-    : "r"(&kUVToRB),   // %4
-      "r"(&kUVToG)     // %5
-    : "cc", "memory", "q0", "q1", "q2", "q3",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
   );
 }
 #endif  // HAS_NV21TORGB565ROW_NEON
@@ -757,29 +675,22 @@ void YUY2ToARGBRow_NEON(const uint8* src_yuy2,
                         uint8* dst_argb,
                         int width) {
   asm volatile (
-    MEMACCESS(3)
-    "vld1.8     {d24}, [%3]                    \n"
-    MEMACCESS(4)
-    "vld1.8     {d25}, [%4]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
-    ".p2align   2                              \n"
+    YUV422TORGB_SETUP_REG
   "1:                                          \n"
     READYUY2
-    YUV422TORGB
+    YUV422TORGB(v22, v21, v20)
     "subs       %2, %2, #8                     \n"
-    "vmov.u8    d23, #255                      \n"
+    "movi       v23.8b, #255                   \n"
     MEMACCESS(1)
-    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
-    "bgt        1b                             \n"
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32      \n"
+    "b.gt       1b                             \n"
     : "+r"(src_yuy2),  // %0
       "+r"(dst_argb),  // %1
       "+r"(width)      // %2
-    : "r"(&kUVToRB),   // %3
-      "r"(&kUVToG)     // %4
-    : "cc", "memory", "q0", "q1", "q2", "q3",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
   );
 }
 #endif  // HAS_YUY2TOARGBROW_NEON
@@ -789,29 +700,22 @@ void UYVYToARGBRow_NEON(const uint8* src_uyvy,
                         uint8* dst_argb,
                         int width) {
   asm volatile (
-    MEMACCESS(3)
-    "vld1.8     {d24}, [%3]                    \n"
-    MEMACCESS(4)
-    "vld1.8     {d25}, [%4]                    \n"
-    "vmov.u8    d26, #128                      \n"
-    "vmov.u16   q14, #74                       \n"
-    "vmov.u16   q15, #16                       \n"
-    ".p2align   2                              \n"
+    YUV422TORGB_SETUP_REG
   "1:                                          \n"
     READUYVY
-    YUV422TORGB
+    YUV422TORGB(v22, v21, v20)
     "subs       %2, %2, #8                     \n"
-    "vmov.u8    d23, #255                      \n"
+    "movi       v23.8b, #255                   \n"
     MEMACCESS(1)
-    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
-    "bgt        1b                             \n"
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], 32      \n"
+    "b.gt       1b                             \n"
     : "+r"(src_uyvy),  // %0
       "+r"(dst_argb),  // %1
       "+r"(width)      // %2
-    : "r"(&kUVToRB),   // %3
-      "r"(&kUVToG)     // %4
-    : "cc", "memory", "q0", "q1", "q2", "q3",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
   );
 }
 #endif  // HAS_UYVYTOARGBROW_NEON
@@ -821,16 +725,15 @@ void UYVYToARGBRow_NEON(const uint8* src_uyvy,
 void SplitUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
                      int width) {
   asm volatile (
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "ld2        {v0.16b, v1.16b}, [%0], #32    \n"  // load 16 pairs of UV
+    "ld2        {v0.16b,v1.16b}, [%0], #32     \n"  // load 16 pairs of UV
     "subs       %3, %3, #16                    \n"  // 16 processed per loop
     MEMACCESS(1)
     "st1        {v0.16b}, [%1], #16            \n"  // store U
     MEMACCESS(2)
     "st1        {v1.16b}, [%2], #16            \n"  // store V
-    "bgt        1b                             \n"
+    "b.gt       1b                             \n"
     : "+r"(src_uv),  // %0
       "+r"(dst_u),   // %1
       "+r"(dst_v),   // %2
@@ -846,7 +749,6 @@ void SplitUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
 void MergeUVRow_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
                      int width) {
   asm volatile (
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
     "ld1        {v0.16b}, [%0], #16            \n"  // load U
@@ -854,8 +756,8 @@ void MergeUVRow_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
     "ld1        {v1.16b}, [%1], #16            \n"  // load V
     "subs       %3, %3, #16                    \n"  // 16 processed per loop
     MEMACCESS(2)
-    "st2        {v0.16b, v1.16b}, [%2], #32    \n"  // store 16 pairs of UV
-    "bgt        1b                             \n"
+    "st2        {v0.16b,v1.16b}, [%2], #32     \n"  // store 16 pairs of UV
+    "b.gt       1b                             \n"
     :
       "+r"(src_u),   // %0
       "+r"(src_v),   // %1
@@ -871,14 +773,13 @@ void MergeUVRow_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
 #ifdef HAS_COPYROW_NEON
 void CopyRow_NEON(const uint8* src, uint8* dst, int count) {
   asm volatile (
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "ld1        {v0.8b-v3.8b}, [%0], #32       \n"  // load 32
+    "ld1        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32       \n"  // load 32
     "subs       %2, %2, #32                    \n"  // 32 processed per loop
     MEMACCESS(1)
-    "st1        {v0.8b-v3.8b}, [%1], #32       \n"  // store 32
-    "bgt        1b                             \n"
+    "st1        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32       \n"  // store 32
+    "b.gt       1b                             \n"
   : "+r"(src),   // %0
     "+r"(dst),   // %1
     "+r"(count)  // %2  // Output registers
@@ -888,35 +789,36 @@ void CopyRow_NEON(const uint8* src, uint8* dst, int count) {
 }
 #endif  // HAS_COPYROW_NEON
 
-// SetRow8 writes 'count' bytes using a 32 bit value repeated.
-#ifdef HAS_SETROW_NEON
-void SetRow_NEON(uint8* dst, uint32 v32, int count) {
+// SetRow writes 'count' bytes using an 8 bit value repeated.
+void SetRow_NEON(uint8* dst, uint8 v8, int count) {
   asm volatile (
-    "dup        v0.4s, %w2                     \n"  // duplicate 4 ints
-    "1:                                        \n"
+    "dup        v0.16b, %w2                    \n"  // duplicate 16 bytes
+  "1:                                          \n"
     "subs      %1, %1, #16                     \n"  // 16 bytes per loop
     MEMACCESS(0)
     "st1        {v0.16b}, [%0], #16            \n"  // store
-    "bgt       1b                              \n"
+    "b.gt      1b                              \n"
   : "+r"(dst),   // %0
     "+r"(count)  // %1
-  : "r"(v32)     // %2
+  : "r"(v8)      // %2
   : "cc", "memory", "v0"
   );
 }
-#endif  // HAS_SETROW_NEON
 
-// TODO(fbarchard): Make fully assembler
-// SetRow32 writes 'count' words using a 32 bit value repeated.
-#ifdef HAS_ARGBSETROWS_NEON
-void ARGBSetRows_NEON(uint8* dst, uint32 v32, int width,
-                      int dst_stride, int height) {
-  for (int y = 0; y < height; ++y) {
-    SetRow_NEON(dst, v32, width << 2);
-    dst += dst_stride;
-  }
+void ARGBSetRow_NEON(uint8* dst, uint32 v32, int count) {
+  asm volatile (
+    "dup        v0.4s, %w2                     \n"  // duplicate 4 ints
+  "1:                                          \n"
+    "subs      %1, %1, #4                      \n"  // 4 ints per loop
+    MEMACCESS(0)
+    "st1        {v0.16b}, [%0], #16            \n"  // store
+    "b.gt      1b                              \n"
+  : "+r"(dst),   // %0
+    "+r"(count)  // %1
+  : "r"(v32)     // %2
+  : "cc", "memory", "v0"
+  );
 }
-#endif  // HAS_ARGBSETROWS_NEON
 
 #ifdef HAS_MIRRORROW_NEON
 void MirrorRow_NEON(const uint8* src, uint8* dst, int width) {
@@ -925,7 +827,6 @@ void MirrorRow_NEON(const uint8* src, uint8* dst, int width) {
     "add        %0, %0, %2                     \n"
     "sub        %0, %0, #16                    \n"
 
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
     "ld1        {v0.16b}, [%0], %3             \n"  // src -= 16
@@ -935,7 +836,7 @@ void MirrorRow_NEON(const uint8* src, uint8* dst, int width) {
     "st1        {v0.D}[1], [%1], #8            \n"  // dst += 16
     MEMACCESS(1)
     "st1        {v0.D}[0], [%1], #8            \n"
-    "bgt        1b                             \n"
+    "b.gt       1b                             \n"
   : "+r"(src),   // %0
     "+r"(dst),   // %1
     "+r"(width)  // %2
@@ -953,7 +854,6 @@ void MirrorUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
     "add        %0, %0, %3, lsl #1             \n"
     "sub        %0, %0, #16                    \n"
 
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
     "ld2        {v0.8b, v1.8b}, [%0], %4       \n"  // src -= 16
@@ -961,10 +861,10 @@ void MirrorUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
     "rev64      v0.8b, v0.8b                   \n"
     "rev64      v1.8b, v1.8b                   \n"
     MEMACCESS(1)
-    "st1        {v0.8b}, [%1], #8               \n"  // dst += 8
+    "st1        {v0.8b}, [%1], #8              \n"  // dst += 8
     MEMACCESS(2)
-    "st1        {v1.8b}, [%2], #8               \n"
-    "bgt        1b                             \n"
+    "st1        {v1.8b}, [%2], #8              \n"
+    "b.gt       1b                             \n"
   : "+r"(src_uv),  // %0
     "+r"(dst_u),   // %1
     "+r"(dst_v),   // %2
@@ -982,7 +882,6 @@ void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width) {
     "add        %0, %0, %2, lsl #2             \n"
     "sub        %0, %0, #16                    \n"
 
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
     "ld1        {v0.16b}, [%0], %3             \n"  // src -= 16
@@ -992,7 +891,7 @@ void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width) {
     "st1        {v0.D}[1], [%1], #8            \n"  // dst += 16
     MEMACCESS(1)
     "st1        {v0.D}[0], [%1], #8            \n"
-    "bgt        1b                             \n"
+    "b.gt       1b                             \n"
   : "+r"(src),   // %0
     "+r"(dst),   // %1
     "+r"(width)  // %2
@@ -1006,14 +905,13 @@ void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width) {
 void RGB24ToARGBRow_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix) {
   asm volatile (
     "movi       v4.8b, #255                    \n"  // Alpha
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "ld3        {v1.8b-v3.8b}, [%0], #24       \n"  // load 8 pixels of RGB24.
+    "ld3        {v1.8b,v2.8b,v3.8b}, [%0], #24 \n"  // load 8 pixels of RGB24.
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
     MEMACCESS(1)
-    "st4        {v1.8b-v4.8b}, [%1], #32       \n"  // store 8 pixels of ARGB.
-    "bgt        1b                             \n"
+    "st4        {v1.8b,v2.8b,v3.8b,v4.8b}, [%1], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
   : "+r"(src_rgb24),  // %0
     "+r"(dst_argb),   // %1
     "+r"(pix)         // %2
@@ -1027,16 +925,15 @@ void RGB24ToARGBRow_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix) {
 void RAWToARGBRow_NEON(const uint8* src_raw, uint8* dst_argb, int pix) {
   asm volatile (
     "movi       v5.8b, #255                    \n"  // Alpha
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "ld3        {v0.8b-v2.8b}, [%0], #24       \n"  // read r g b
+    "ld3        {v0.8b,v1.8b,v2.8b}, [%0], #24 \n"  // read r g b
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "mov        v3.8b, v1.8b                   \n"  // move g
-    "mov        v4.8b, v0.8b                   \n"  // move r
+    "orr        v3.8b, v1.8b, v1.8b            \n"  // move g
+    "orr        v4.8b, v0.8b, v0.8b            \n"  // move r
     MEMACCESS(1)
-    "st4        {v2.8b-v5.8b}, [%1], #32       \n"  // store b g r a
-    "bgt        1b                             \n"
+    "st4        {v2.8b,v3.8b,v4.8b,v5.8b}, [%1], #32 \n"  // store b g r a
+    "b.gt       1b                             \n"
   : "+r"(src_raw),   // %0
     "+r"(dst_argb),  // %1
     "+r"(pix)        // %2
@@ -1047,118 +944,127 @@ void RAWToARGBRow_NEON(const uint8* src_raw, uint8* dst_argb, int pix) {
 #endif  // HAS_RAWTOARGBROW_NEON
 
 #define RGB565TOARGB                                                           \
-    "vshrn.u16  d6, q0, #5                     \n"  /* G xxGGGGGG           */ \
-    "vuzp.u8    d0, d1                         \n"  /* d0 xxxBBBBB RRRRRxxx */ \
-    "vshl.u8    d6, d6, #2                     \n"  /* G GGGGGG00 upper 6   */ \
-    "vshr.u8    d1, d1, #3                     \n"  /* R 000RRRRR lower 5   */ \
-    "vshl.u8    q0, q0, #3                     \n"  /* B,R BBBBB000 upper 5 */ \
-    "vshr.u8    q2, q0, #5                     \n"  /* B,R 00000BBB lower 3 */ \
-    "vorr.u8    d0, d0, d4                     \n"  /* B                    */ \
-    "vshr.u8    d4, d6, #6                     \n"  /* G 000000GG lower 2   */ \
-    "vorr.u8    d2, d1, d5                     \n"  /* R                    */ \
-    "vorr.u8    d1, d4, d6                     \n"  /* G                    */
+    "shrn       v6.8b, v0.8h, #5               \n"  /* G xxGGGGGG           */ \
+    "shl        v6.8b, v6.8b, #2               \n"  /* G GGGGGG00 upper 6   */ \
+    "ushr       v4.8b, v6.8b, #6               \n"  /* G 000000GG lower 2   */ \
+    "orr        v1.8b, v4.8b, v6.8b            \n"  /* G                    */ \
+    "xtn        v2.8b, v0.8h                   \n"  /* B xxxBBBBB           */ \
+    "ushr       v0.8h, v0.8h, #11              \n"  /* R 000RRRRR           */ \
+    "xtn2       v2.16b,v0.8h                   \n"  /* R in upper part      */ \
+    "shl        v2.16b, v2.16b, #3             \n"  /* R,B BBBBB000 upper 5 */ \
+    "ushr       v0.16b, v2.16b, #5             \n"  /* R,B 00000BBB lower 3 */ \
+    "orr        v0.16b, v0.16b, v2.16b         \n"  /* R,B                  */ \
+    "dup        v2.2D, v0.D[1]                 \n"  /* R                    */
 
 #ifdef HAS_RGB565TOARGBROW_NEON
 void RGB565ToARGBRow_NEON(const uint8* src_rgb565, uint8* dst_argb, int pix) {
   asm volatile (
-    "vmov.u8    d3, #255                       \n"  // Alpha
-    ".p2align   2                              \n"
+    "movi       v3.8b, #255                    \n"  // Alpha
   "1:                                          \n"
     MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 RGB565 pixels.
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
     RGB565TOARGB
     MEMACCESS(1)
-    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
-    "bgt        1b                             \n"
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
   : "+r"(src_rgb565),  // %0
     "+r"(dst_argb),    // %1
     "+r"(pix)          // %2
   :
-  : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v6"  // Clobber List
   );
 }
 #endif  // HAS_RGB565TOARGBROW_NEON
 
 #define ARGB1555TOARGB                                                         \
-    "vshrn.u16  d7, q0, #8                     \n"  /* A Arrrrrxx           */ \
-    "vshr.u8    d6, d7, #2                     \n"  /* R xxxRRRRR           */ \
-    "vshrn.u16  d5, q0, #5                     \n"  /* G xxxGGGGG           */ \
-    "vmovn.u16  d4, q0                         \n"  /* B xxxBBBBB           */ \
-    "vshr.u8    d7, d7, #7                     \n"  /* A 0000000A           */ \
-    "vneg.s8    d7, d7                         \n"  /* A AAAAAAAA upper 8   */ \
-    "vshl.u8    d6, d6, #3                     \n"  /* R RRRRR000 upper 5   */ \
-    "vshr.u8    q1, q3, #5                     \n"  /* R,A 00000RRR lower 3 */ \
-    "vshl.u8    q0, q2, #3                     \n"  /* B,G BBBBB000 upper 5 */ \
-    "vshr.u8    q2, q0, #5                     \n"  /* B,G 00000BBB lower 3 */ \
-    "vorr.u8    q1, q1, q3                     \n"  /* R,A                  */ \
-    "vorr.u8    q0, q0, q2                     \n"  /* B,G                  */ \
+    "ushr       v2.8h, v0.8h, #10              \n"  /* R xxxRRRRR           */ \
+    "shl        v2.8h, v2.8h, #3               \n"  /* R RRRRR000 upper 5   */ \
+    "xtn        v3.8b, v2.8h                   \n"  /* RRRRR000 AAAAAAAA    */ \
+                                                                               \
+    "sshr       v2.8h, v0.8h, #15              \n"  /* A AAAAAAAA           */ \
+    "xtn2       v3.16b, v2.8h                  \n"                             \
+                                                                               \
+    "xtn        v2.8b, v0.8h                   \n"  /* B xxxBBBBB           */ \
+    "shrn2      v2.16b,v0.8h, #5               \n"  /* G xxxGGGGG           */ \
+                                                                               \
+    "ushr       v1.16b, v3.16b, #5             \n"  /* R,A 00000RRR lower 3 */ \
+    "shl        v0.16b, v2.16b, #3             \n"  /* B,G BBBBB000 upper 5 */ \
+    "ushr       v2.16b, v0.16b, #5             \n"  /* B,G 00000BBB lower 3 */ \
+                                                                               \
+    "orr        v0.16b, v0.16b, v2.16b         \n"  /* B,G                  */ \
+    "orr        v2.16b, v1.16b, v3.16b         \n"  /* R,A                  */ \
+    "dup        v1.2D, v0.D[1]                 \n"                             \
+    "dup        v3.2D, v2.D[1]                 \n"
 
 // RGB555TOARGB is same as ARGB1555TOARGB but ignores alpha.
 #define RGB555TOARGB                                                           \
-    "vshrn.u16  d6, q0, #5                     \n"  /* G xxxGGGGG           */ \
-    "vuzp.u8    d0, d1                         \n"  /* d0 xxxBBBBB xRRRRRxx */ \
-    "vshl.u8    d6, d6, #3                     \n"  /* G GGGGG000 upper 5   */ \
-    "vshr.u8    d1, d1, #2                     \n"  /* R 00xRRRRR lower 5   */ \
-    "vshl.u8    q0, q0, #3                     \n"  /* B,R BBBBB000 upper 5 */ \
-    "vshr.u8    q2, q0, #5                     \n"  /* B,R 00000BBB lower 3 */ \
-    "vorr.u8    d0, d0, d4                     \n"  /* B                    */ \
-    "vshr.u8    d4, d6, #5                     \n"  /* G 00000GGG lower 3   */ \
-    "vorr.u8    d2, d1, d5                     \n"  /* R                    */ \
-    "vorr.u8    d1, d4, d6                     \n"  /* G                    */
+    "ushr       v2.8h, v0.8h, #10              \n"  /* R xxxRRRRR           */ \
+    "shl        v2.8h, v2.8h, #3               \n"  /* R RRRRR000 upper 5   */ \
+    "xtn        v3.8b, v2.8h                   \n"  /* RRRRR000             */ \
+                                                                               \
+    "xtn        v2.8b, v0.8h                   \n"  /* B xxxBBBBB           */ \
+    "shrn2      v2.16b,v0.8h, #5               \n"  /* G xxxGGGGG           */ \
+                                                                               \
+    "ushr       v1.16b, v3.16b, #5             \n"  /* R   00000RRR lower 3 */ \
+    "shl        v0.16b, v2.16b, #3             \n"  /* B,G BBBBB000 upper 5 */ \
+    "ushr       v2.16b, v0.16b, #5             \n"  /* B,G 00000BBB lower 3 */ \
+                                                                               \
+    "orr        v0.16b, v0.16b, v2.16b         \n"  /* B,G                  */ \
+    "orr        v2.16b, v1.16b, v3.16b         \n"  /* R                    */ \
+    "dup        v1.2D, v0.D[1]                 \n"  /* G */                    \
 
 #ifdef HAS_ARGB1555TOARGBROW_NEON
 void ARGB1555ToARGBRow_NEON(const uint8* src_argb1555, uint8* dst_argb,
                             int pix) {
   asm volatile (
-    "vmov.u8    d3, #255                       \n"  // Alpha
-    ".p2align   2                              \n"
+    "movi       v3.8b, #255                    \n"  // Alpha
   "1:                                          \n"
     MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB1555 pixels.
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
     ARGB1555TOARGB
     MEMACCESS(1)
-    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
-    "bgt        1b                             \n"
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
   : "+r"(src_argb1555),  // %0
     "+r"(dst_argb),    // %1
     "+r"(pix)          // %2
   :
-  : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
   );
 }
 #endif  // HAS_ARGB1555TOARGBROW_NEON
 
 #define ARGB4444TOARGB                                                         \
-    "vuzp.u8    d0, d1                         \n"  /* d0 BG, d1 RA         */ \
-    "vshl.u8    q2, q0, #4                     \n"  /* B,R BBBB0000         */ \
-    "vshr.u8    q1, q0, #4                     \n"  /* G,A 0000GGGG         */ \
-    "vshr.u8    q0, q2, #4                     \n"  /* B,R 0000BBBB         */ \
-    "vorr.u8    q0, q0, q2                     \n"  /* B,R BBBBBBBB         */ \
-    "vshl.u8    q2, q1, #4                     \n"  /* G,A GGGG0000         */ \
-    "vorr.u8    q1, q1, q2                     \n"  /* G,A GGGGGGGG         */ \
-    "vswp.u8    d1, d2                         \n"  /* B,R,G,A -> B,G,R,A   */
+    "shrn       v1.8b,  v0.8h, #8              \n"  /* v1(l) AR             */ \
+    "xtn2       v1.16b, v0.8h                  \n"  /* v1(h) GB             */ \
+    "shl        v2.16b, v1.16b, #4             \n"  /* B,R BBBB0000         */ \
+    "ushr       v3.16b, v1.16b, #4             \n"  /* G,A 0000GGGG         */ \
+    "ushr       v0.16b, v2.16b, #4             \n"  /* B,R 0000BBBB         */ \
+    "shl        v1.16b, v3.16b, #4             \n"  /* G,A GGGG0000         */ \
+    "orr        v2.16b, v0.16b, v2.16b         \n"  /* B,R BBBBBBBB         */ \
+    "orr        v3.16b, v1.16b, v3.16b         \n"  /* G,A GGGGGGGG         */ \
+    "dup        v0.2D, v2.D[1]                 \n"                             \
+    "dup        v1.2D, v3.D[1]                 \n"
 
 #ifdef HAS_ARGB4444TOARGBROW_NEON
 void ARGB4444ToARGBRow_NEON(const uint8* src_argb4444, uint8* dst_argb,
                             int pix) {
   asm volatile (
-    "vmov.u8    d3, #255                       \n"  // Alpha
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB4444 pixels.
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
     ARGB4444TOARGB
     MEMACCESS(1)
-    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
-    "bgt        1b                             \n"
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
   : "+r"(src_argb4444),  // %0
     "+r"(dst_argb),    // %1
     "+r"(pix)          // %2
   :
-  : "cc", "memory", "q0", "q1", "q2"  // Clobber List
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4"  // Clobber List
   );
 }
 #endif  // HAS_ARGB4444TOARGBROW_NEON
@@ -1166,14 +1072,13 @@ void ARGB4444ToARGBRow_NEON(const uint8* src_argb4444, uint8* dst_argb,
 #ifdef HAS_ARGBTORGB24ROW_NEON
 void ARGBToRGB24Row_NEON(const uint8* src_argb, uint8* dst_rgb24, int pix) {
   asm volatile (
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "ld4        {v1.8b-v4.8b}, [%0], #32       \n"  // load 8 pixels of ARGB.
+    "ld4        {v1.8b,v2.8b,v3.8b,v4.8b}, [%0], #32 \n"  // load 8 ARGB pixels
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
     MEMACCESS(1)
-    "st3        {v1.8b-v3.8b}, [%1], #24       \n"  // store 8 pixels of RGB24.
-    "bgt        1b                             \n"
+    "st3        {v1.8b,v2.8b,v3.8b}, [%1], #24 \n"  // store 8 pixels of RGB24.
+    "b.gt       1b                             \n"
   : "+r"(src_argb),   // %0
     "+r"(dst_rgb24),  // %1
     "+r"(pix)         // %2
@@ -1186,16 +1091,15 @@ void ARGBToRGB24Row_NEON(const uint8* src_argb, uint8* dst_rgb24, int pix) {
 #ifdef HAS_ARGBTORAWROW_NEON
 void ARGBToRAWRow_NEON(const uint8* src_argb, uint8* dst_raw, int pix) {
   asm volatile (
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "ld4        {v1.8b-v4.8b}, [%0], #32       \n"  // load b g r a
+    "ld4        {v1.8b,v2.8b,v3.8b,v4.8b}, [%0], #32 \n"  // load b g r a
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "mov        v4.8b, v2.8b                   \n"  // mov g
-    "mov        v5.8b, v1.8b                   \n"  // mov b
+    "orr        v4.8b, v2.8b, v2.8b            \n"  // mov g
+    "orr        v5.8b, v1.8b, v1.8b            \n"  // mov b
     MEMACCESS(1)
-    "st3        {v3.8b-v5.8b}, [%1], #24       \n"  // store r g b
-    "bgt        1b                             \n"
+    "st3        {v3.8b,v4.8b,v5.8b}, [%1], #24 \n"  // store r g b
+    "b.gt       1b                             \n"
   : "+r"(src_argb),  // %0
     "+r"(dst_raw),   // %1
     "+r"(pix)        // %2
@@ -1208,14 +1112,13 @@ void ARGBToRAWRow_NEON(const uint8* src_argb, uint8* dst_raw, int pix) {
 #ifdef HAS_YUY2TOYROW_NEON
 void YUY2ToYRow_NEON(const uint8* src_yuy2, uint8* dst_y, int pix) {
   asm volatile (
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "ld2        {v0.16b, v1.16b}, [%0], #32    \n"  // load 16 pixels of YUY2.
+    "ld2        {v0.16b,v1.16b}, [%0], #32     \n"  // load 16 pixels of YUY2.
     "subs       %2, %2, #16                    \n"  // 16 processed per loop.
     MEMACCESS(1)
     "st1        {v0.16b}, [%1], #16            \n"  // store 16 pixels of Y.
-    "bgt        1b                             \n"
+    "b.gt       1b                             \n"
   : "+r"(src_yuy2),  // %0
     "+r"(dst_y),     // %1
     "+r"(pix)        // %2
@@ -1228,14 +1131,13 @@ void YUY2ToYRow_NEON(const uint8* src_yuy2, uint8* dst_y, int pix) {
 #ifdef HAS_UYVYTOYROW_NEON
 void UYVYToYRow_NEON(const uint8* src_uyvy, uint8* dst_y, int pix) {
   asm volatile (
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "ld2        {v0.16b, v1.16b}, [%0], #32    \n"  // load 16 pixels of UYVY.
+    "ld2        {v0.16b,v1.16b}, [%0], #32     \n"  // load 16 pixels of UYVY.
     "subs       %2, %2, #16                    \n"  // 16 processed per loop.
     MEMACCESS(1)
     "st1        {v1.16b}, [%1], #16            \n"  // store 16 pixels of Y.
-    "bgt        1b                             \n"
+    "b.gt       1b                             \n"
   : "+r"(src_uyvy),  // %0
     "+r"(dst_y),     // %1
     "+r"(pix)        // %2
@@ -1249,16 +1151,15 @@ void UYVYToYRow_NEON(const uint8* src_uyvy, uint8* dst_y, int pix) {
 void YUY2ToUV422Row_NEON(const uint8* src_yuy2, uint8* dst_u, uint8* dst_v,
                          int pix) {
   asm volatile (
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "ld4        {v0.8b-v3.8b}, [%0], #32       \n"  // load 16 pixels of YUY2.
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 YUY2 pixels
     "subs       %3, %3, #16                    \n"  // 16 pixels = 8 UVs.
     MEMACCESS(1)
     "st1        {v1.8b}, [%1], #8              \n"  // store 8 U.
     MEMACCESS(2)
     "st1        {v3.8b}, [%2], #8              \n"  // store 8 V.
-    "bgt        1b                             \n"
+    "b.gt       1b                             \n"
   : "+r"(src_yuy2),  // %0
     "+r"(dst_u),     // %1
     "+r"(dst_v),     // %2
@@ -1273,16 +1174,15 @@ void YUY2ToUV422Row_NEON(const uint8* src_yuy2, uint8* dst_u, uint8* dst_v,
 void UYVYToUV422Row_NEON(const uint8* src_uyvy, uint8* dst_u, uint8* dst_v,
                          int pix) {
   asm volatile (
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "ld4        {v0.8b-v3.8b}, [%0], #32       \n"  // load 16 pixels of UYVY.
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 UYVY pixels
     "subs       %3, %3, #16                    \n"  // 16 pixels = 8 UVs.
     MEMACCESS(1)
     "st1        {v0.8b}, [%1], #8              \n"  // store 8 U.
     MEMACCESS(2)
     "st1        {v2.8b}, [%2], #8              \n"  // store 8 V.
-    "bgt        1b                             \n"
+    "b.gt       1b                             \n"
   : "+r"(src_uyvy),  // %0
     "+r"(dst_u),     // %1
     "+r"(dst_v),     // %2
@@ -1296,29 +1196,29 @@ void UYVYToUV422Row_NEON(const uint8* src_uyvy, uint8* dst_u, uint8* dst_v,
 #ifdef HAS_YUY2TOUVROW_NEON
 void YUY2ToUVRow_NEON(const uint8* src_yuy2, int stride_yuy2,
                       uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_yuy2b = src_yuy2 + stride_yuy2;
   asm volatile (
-    "add        %x1, %x0, %w1, sxtw            \n"  // stride + src_yuy2
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "ld4        {v0.8b-v3.8b}, [%0], #32       \n"  // load 16 pixels of YUY2.
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 pixels
     "subs       %4, %4, #16                    \n"  // 16 pixels = 8 UVs.
     MEMACCESS(1)
-    "ld4        {v4.8b-v7.8b}, [%1], #32       \n"  // load next row YUY2.
+    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load next row
     "urhadd     v1.8b, v1.8b, v5.8b            \n"  // average rows of U
     "urhadd     v3.8b, v3.8b, v7.8b            \n"  // average rows of V
     MEMACCESS(2)
     "st1        {v1.8b}, [%2], #8              \n"  // store 8 U.
     MEMACCESS(3)
     "st1        {v3.8b}, [%3], #8              \n"  // store 8 V.
-    "bgt        1b                             \n"
+    "b.gt       1b                             \n"
   : "+r"(src_yuy2),     // %0
-    "+r"(stride_yuy2),  // %1
+    "+r"(src_yuy2b),    // %1
     "+r"(dst_u),        // %2
     "+r"(dst_v),        // %3
     "+r"(pix)           // %4
   :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"  // Clobber List
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4",
+    "v5", "v6", "v7"  // Clobber List
   );
 }
 #endif  // HAS_YUY2TOUVROW_NEON
@@ -1326,84 +1226,33 @@ void YUY2ToUVRow_NEON(const uint8* src_yuy2, int stride_yuy2,
 #ifdef HAS_UYVYTOUVROW_NEON
 void UYVYToUVRow_NEON(const uint8* src_uyvy, int stride_uyvy,
                       uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_uyvyb = src_uyvy + stride_uyvy;
   asm volatile (
-    "add        %x1, %x0, %w1, sxtw            \n"  // stride + src_uyvy
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "ld4        {v0.8b-v3.8b}, [%0], #32       \n"  // load 16 pixels of UYVY.
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 pixels
     "subs       %4, %4, #16                    \n"  // 16 pixels = 8 UVs.
     MEMACCESS(1)
-    "ld4        {v4.8b-v7.8b}, [%1], #32       \n"  // load next row UYVY.
+    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load next row
     "urhadd     v0.8b, v0.8b, v4.8b            \n"  // average rows of U
     "urhadd     v2.8b, v2.8b, v6.8b            \n"  // average rows of V
     MEMACCESS(2)
     "st1        {v0.8b}, [%2], #8              \n"  // store 8 U.
     MEMACCESS(3)
     "st1        {v2.8b}, [%3], #8              \n"  // store 8 V.
-    "bgt        1b                             \n"
+    "b.gt       1b                             \n"
   : "+r"(src_uyvy),     // %0
-    "+r"(stride_uyvy),  // %1
+    "+r"(src_uyvyb),    // %1
     "+r"(dst_u),        // %2
     "+r"(dst_v),        // %3
     "+r"(pix)           // %4
   :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"  // Clobber List
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4",
+    "v5", "v6", "v7"  // Clobber List
   );
 }
 #endif  // HAS_UYVYTOUVROW_NEON
 
-#ifdef HAS_HALFROW_NEON
-void HalfRow_NEON(const uint8* src_uv, int src_uv_stride,
-                  uint8* dst_uv, int pix) {
-  asm volatile (
-    // change the stride to row 2 pointer
-    "add        %x1, %x0, %w1, sxtw            \n"
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.16b}, [%0], #16            \n"  // load row 1 16 pixels.
-    "subs       %3, %3, #16                    \n"  // 16 processed per loop
-    MEMACCESS(1)
-    "ld1        {v1.16b}, [%1], #16            \n"  // load row 2 16 pixels.
-    "urhadd     v0.16b, v0.16b, v1.16b         \n"  // average row 1 and 2
-    MEMACCESS(2)
-    "st1        {v0.16b}, [%2], #16            \n"
-    "bgt        1b                             \n"
-  : "+r"(src_uv),         // %0
-    "+r"(src_uv_stride),  // %1
-    "+r"(dst_uv),         // %2
-    "+r"(pix)             // %3
-  :
-  : "cc", "memory", "v0", "v1"  // Clobber List
-  );
-}
-#endif  // HAS_HALFROW_NEON
-
-// Select 2 channels from ARGB on alternating pixels.  e.g.  BGBGBGBG
-#ifdef HAS_ARGBTOBAYERROW_NEON
-void ARGBToBayerRow_NEON(const uint8* src_argb, uint8* dst_bayer,
-                         uint32 selector, int pix) {
-  asm volatile (
-    "mov        v2.s[0], %w3                   \n"  // selector
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.16b, v1.16b}, [%0], 32     \n"  // load row 8 pixels.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop
-    "tbl        v4.8b, {v0.16b}, v2.8b         \n"  // look up 4 pixels
-    "tbl        v5.8b, {v1.16b}, v2.8b         \n"  // look up 4 pixels
-    "trn1       v4.4s, v4.4s, v5.4s            \n"  // combine 8 pixels
-    MEMACCESS(1)
-    "st1        {v4.8b}, [%1], #8              \n"  // store 8.
-    "bgt        1b                             \n"
-  : "+r"(src_argb),   // %0
-    "+r"(dst_bayer),  // %1
-    "+r"(pix)         // %2
-  : "r"(selector)     // %3
-  : "cc", "memory", "v0", "v1", "v2", "v4", "v5"   // Clobber List
-  );
-}
-#endif  // HAS_ARGBTOBAYERROW_NEON
-
 // Select G channels from ARGB.  e.g.  GGGGGGGG
 #ifdef HAS_ARGBTOBAYERGGROW_NEON
 void ARGBToBayerGGRow_NEON(const uint8* src_argb, uint8* dst_bayer,
@@ -1411,11 +1260,11 @@ void ARGBToBayerGGRow_NEON(const uint8* src_argb, uint8* dst_bayer,
   asm volatile (
   "1:                                          \n"
     MEMACCESS(0)
-    "ld4        {v0.8b-v3.8b}, [%0], #32       \n"  // load row 8 pixels.
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load row 8 pixels
     "subs       %2, %2, #8                     \n"  // 8 processed per loop
     MEMACCESS(1)
     "st1        {v1.8b}, [%1], #8              \n"  // store 8 G's.
-    "bgt        1b                             \n"
+    "b.gt       1b                             \n"
   : "+r"(src_argb),   // %0
     "+r"(dst_bayer),  // %1
     "+r"(pix)         // %2
@@ -1439,7 +1288,7 @@ void ARGBShuffleRow_NEON(const uint8* src_argb, uint8* dst_argb,
     "tbl        v1.16b, {v0.16b}, v2.16b       \n"  // look up 4 pixels
     MEMACCESS(1)
     "st1        {v1.16b}, [%1], #16            \n"  // store 4.
-    "bgt        1b                             \n"
+    "b.gt       1b                             \n"
   : "+r"(src_argb),  // %0
     "+r"(dst_argb),  // %1
     "+r"(pix)        // %2
@@ -1455,19 +1304,18 @@ void I422ToYUY2Row_NEON(const uint8* src_y,
                         const uint8* src_v,
                         uint8* dst_yuy2, int width) {
   asm volatile (
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
     "ld2        {v0.8b, v1.8b}, [%0], #16      \n"  // load 16 Ys
-    "mov        v2.8b, v1.8b                   \n"
+    "orr        v2.8b, v1.8b, v1.8b            \n"
     MEMACCESS(1)
     "ld1        {v1.8b}, [%1], #8              \n"  // load 8 Us
     MEMACCESS(2)
     "ld1        {v3.8b}, [%2], #8              \n"  // load 8 Vs
     "subs       %4, %4, #16                    \n"  // 16 pixels
     MEMACCESS(3)
-    "st4        {v0.8b-v3.8b}, [%3], #32       \n"  // Store 8 YUY2/16 pixels.
-    "bgt        1b                             \n"
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%3], #32 \n"  // Store 16 pixels.
+    "b.gt       1b                             \n"
   : "+r"(src_y),     // %0
     "+r"(src_u),     // %1
     "+r"(src_v),     // %2
@@ -1485,19 +1333,18 @@ void I422ToUYVYRow_NEON(const uint8* src_y,
                         const uint8* src_v,
                         uint8* dst_uyvy, int width) {
   asm volatile (
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "ld2        {v1.8b, v2.8b}, [%0], #16      \n"  // load 16 Ys
-    "mov        v3.8b, v2.8b                   \n"
+    "ld2        {v1.8b,v2.8b}, [%0], #16       \n"  // load 16 Ys
+    "orr        v3.8b, v2.8b, v2.8b            \n"
     MEMACCESS(1)
     "ld1        {v0.8b}, [%1], #8              \n"  // load 8 Us
     MEMACCESS(2)
     "ld1        {v2.8b}, [%2], #8              \n"  // load 8 Vs
     "subs       %4, %4, #16                    \n"  // 16 pixels
     MEMACCESS(3)
-    "st4        {v0.8b-v3.8b}, [%3], #32       \n"  // Store 8 UYVY/16 pixels.
-    "bgt        1b                             \n"
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%3], #32 \n"  // Store 16 pixels.
+    "b.gt       1b                             \n"
   : "+r"(src_y),     // %0
     "+r"(src_u),     // %1
     "+r"(src_v),     // %2
@@ -1512,20 +1359,19 @@ void I422ToUYVYRow_NEON(const uint8* src_y,
 #ifdef HAS_ARGBTORGB565ROW_NEON
 void ARGBToRGB565Row_NEON(const uint8* src_argb, uint8* dst_rgb565, int pix) {
   asm volatile (
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "ld4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%0], #32 \n"  // load 8 pixels
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
     ARGBTORGB565
     MEMACCESS(1)
-    "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels RGB565.
-    "bgt        1b                             \n"
+    "st1        {v0.16b}, [%1], #16            \n"  // store 8 pixels RGB565.
+    "b.gt       1b                             \n"
   : "+r"(src_argb),  // %0
     "+r"(dst_rgb565),  // %1
     "+r"(pix)        // %2
   :
-  : "cc", "memory", "q0", "q8", "q9", "q10", "q11"
+  : "cc", "memory", "v0", "v20", "v21", "v22", "v23"
   );
 }
 #endif  // HAS_ARGBTORGB565ROW_NEON
@@ -1534,20 +1380,19 @@ void ARGBToRGB565Row_NEON(const uint8* src_argb, uint8* dst_rgb565, int pix) {
 void ARGBToARGB1555Row_NEON(const uint8* src_argb, uint8* dst_argb1555,
                             int pix) {
   asm volatile (
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "ld4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%0], #32 \n"  // load 8 pixels
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
     ARGBTOARGB1555
     MEMACCESS(1)
-    "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels ARGB1555.
-    "bgt        1b                             \n"
+    "st1        {v0.16b}, [%1], #16            \n"  // store 8 pixels ARGB1555.
+    "b.gt       1b                             \n"
   : "+r"(src_argb),  // %0
     "+r"(dst_argb1555),  // %1
     "+r"(pix)        // %2
   :
-  : "cc", "memory", "q0", "q8", "q9", "q10", "q11"
+  : "cc", "memory", "v0", "v20", "v21", "v22", "v23"
   );
 }
 #endif  // HAS_ARGBTOARGB1555ROW_NEON
@@ -1556,21 +1401,20 @@ void ARGBToARGB1555Row_NEON(const uint8* src_argb, uint8* dst_argb1555,
 void ARGBToARGB4444Row_NEON(const uint8* src_argb, uint8* dst_argb4444,
                             int pix) {
   asm volatile (
-    "vmov.u8    d4, #0x0f                      \n"  // bits to clear with vbic.
-    ".p2align   2                              \n"
+    "movi       v4.16b, #0x0f                  \n"  // bits to clear with vbic.
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "ld4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%0], #32 \n"  // load 8 pixels
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
     ARGBTOARGB4444
     MEMACCESS(1)
-    "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels ARGB4444.
-    "bgt        1b                             \n"
+    "st1        {v0.16b}, [%1], #16            \n"  // store 8 pixels ARGB4444.
+    "b.gt       1b                             \n"
   : "+r"(src_argb),      // %0
     "+r"(dst_argb4444),  // %1
     "+r"(pix)            // %2
   :
-  : "cc", "memory", "q0", "q8", "q9", "q10", "q11"
+  : "cc", "memory", "v0", "v1", "v4", "v20", "v21", "v22", "v23"
   );
 }
 #endif  // HAS_ARGBTOARGB4444ROW_NEON
@@ -1582,10 +1426,9 @@ void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) {
     "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
     "movi       v6.8b, #33                     \n"  // R * 0.2578 coefficient
     "movi       v7.8b, #16                     \n"  // Add 16 constant
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "ld4        {v0.8b-v3.8b}, [%0], #32       \n"  // load 8 ARGB pixels.
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
     "umull      v3.8h, v0.8b, v4.8b            \n"  // B
     "umlal      v3.8h, v1.8b, v5.8b            \n"  // G
@@ -1594,7 +1437,7 @@ void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) {
     "uqadd      v0.8b, v0.8b, v7.8b            \n"
     MEMACCESS(1)
     "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
-    "bgt        1b                             \n"
+    "b.gt       1b                             \n"
   : "+r"(src_argb),  // %0
     "+r"(dst_y),     // %1
     "+r"(pix)        // %2
@@ -1610,10 +1453,9 @@ void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) {
     "movi       v4.8b, #15                     \n"  // B * 0.11400 coefficient
     "movi       v5.8b, #75                     \n"  // G * 0.58700 coefficient
     "movi       v6.8b, #38                     \n"  // R * 0.29900 coefficient
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "ld4        {v0.8b-v3.8b}, [%0], #32       \n"  // load 8 ARGB pixels.
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
     "umull      v3.8h, v0.8b, v4.8b            \n"  // B
     "umlal      v3.8h, v1.8b, v5.8b            \n"  // G
@@ -1621,7 +1463,7 @@ void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) {
     "sqrshrun   v0.8b, v3.8h, #7               \n"  // 15 bit to 8 bit Y
     MEMACCESS(1)
     "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
-    "bgt        1b                             \n"
+    "b.gt       1b                             \n"
   : "+r"(src_argb),  // %0
     "+r"(dst_y),     // %1
     "+r"(pix)        // %2
@@ -1636,41 +1478,41 @@ void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) {
 void ARGBToUV444Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
                          int pix) {
   asm volatile (
-    "vmov.u8    d24, #112                      \n"  // UB / VR 0.875 coefficient
-    "vmov.u8    d25, #74                       \n"  // UG -0.5781 coefficient
-    "vmov.u8    d26, #38                       \n"  // UR -0.2969 coefficient
-    "vmov.u8    d27, #18                       \n"  // VB -0.1406 coefficient
-    "vmov.u8    d28, #94                       \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-    ".p2align   2                              \n"
+    "movi       v24.8b, #112                   \n"  // UB / VR 0.875 coefficient
+    "movi       v25.8b, #74                    \n"  // UG -0.5781 coefficient
+    "movi       v26.8b, #38                    \n"  // UR -0.2969 coefficient
+    "movi       v27.8b, #18                    \n"  // VB -0.1406 coefficient
+    "movi       v28.8b, #94                    \n"  // VG -0.7344 coefficient
+    "movi       v29.16b,#0x80                  \n"  // 128.5
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
     "subs       %3, %3, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q2, d0, d24                    \n"  // B
-    "vmlsl.u8   q2, d1, d25                    \n"  // G
-    "vmlsl.u8   q2, d2, d26                    \n"  // R
-    "vadd.u16   q2, q2, q15                    \n"  // +128 -> unsigned
+    "umull      v4.8h, v0.8b, v24.8b           \n"  // B
+    "umlsl      v4.8h, v1.8b, v25.8b           \n"  // G
+    "umlsl      v4.8h, v2.8b, v26.8b           \n"  // R
+    "add        v4.8h, v4.8h, v29.8h           \n"  // +128 -> unsigned
 
-    "vmull.u8   q3, d2, d24                    \n"  // R
-    "vmlsl.u8   q3, d1, d28                    \n"  // G
-    "vmlsl.u8   q3, d0, d27                    \n"  // B
-    "vadd.u16   q3, q3, q15                    \n"  // +128 -> unsigned
+    "umull      v3.8h, v2.8b, v24.8b           \n"  // R
+    "umlsl      v3.8h, v1.8b, v28.8b           \n"  // G
+    "umlsl      v3.8h, v0.8b, v27.8b           \n"  // B
+    "add        v3.8h, v3.8h, v29.8h           \n"  // +128 -> unsigned
 
-    "vqshrn.u16  d0, q2, #8                    \n"  // 16 bit to 8 bit U
-    "vqshrn.u16  d1, q3, #8                    \n"  // 16 bit to 8 bit V
+    "uqshrn     v0.8b, v4.8h, #8               \n"  // 16 bit to 8 bit U
+    "uqshrn     v1.8b, v3.8h, #8               \n"  // 16 bit to 8 bit V
 
     MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels U.
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels U.
     MEMACCESS(2)
-    "vst1.8     {d1}, [%2]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
+    "st1        {v1.8b}, [%2], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
   : "+r"(src_argb),  // %0
     "+r"(dst_u),     // %1
     "+r"(dst_v),     // %2
     "+r"(pix)        // %3
   :
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q12", "q13", "q14", "q15"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4",
+    "v24", "v25", "v26", "v27", "v28", "v29"
   );
 }
 #endif  // HAS_ARGBTOUV444ROW_NEON
@@ -1680,49 +1522,41 @@ void ARGBToUV444Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
 void ARGBToUV422Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
                          int pix) {
   asm volatile (
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-    ".p2align   2                              \n"
+    RGBTOUV_SETUP_REG
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
-    MEMACCESS(0)
-    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
 
-    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
 
     "subs       %3, %3, #16                    \n"  // 16 processed per loop.
-    "vmul.s16   q8, q0, q10                    \n"  // B
-    "vmls.s16   q8, q1, q11                    \n"  // G
-    "vmls.s16   q8, q2, q12                    \n"  // R
-    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "mul        v3.8h, v0.8h, v20.8h           \n"  // B
+    "mls        v3.8h, v1.8h, v21.8h           \n"  // G
+    "mls        v3.8h, v2.8h, v22.8h           \n"  // R
+    "add        v3.8h, v3.8h, v25.8h           \n"  // +128 -> unsigned
 
-    "vmul.s16   q9, q2, q10                    \n"  // R
-    "vmls.s16   q9, q1, q14                    \n"  // G
-    "vmls.s16   q9, q0, q13                    \n"  // B
-    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "mul        v4.8h, v2.8h, v20.8h           \n"  // R
+    "mls        v4.8h, v1.8h, v24.8h           \n"  // G
+    "mls        v4.8h, v0.8h, v23.8h           \n"  // B
+    "add        v4.8h, v4.8h, v25.8h           \n"  // +128 -> unsigned
 
-    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
-    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    "uqshrn     v0.8b, v3.8h, #8               \n"  // 16 bit to 8 bit U
+    "uqshrn     v1.8b, v4.8h, #8               \n"  // 16 bit to 8 bit V
 
     MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels U.
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels U.
     MEMACCESS(2)
-    "vst1.8     {d1}, [%2]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
+    "st1        {v1.8b}, [%2], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
   : "+r"(src_argb),  // %0
     "+r"(dst_u),     // %1
     "+r"(dst_v),     // %2
     "+r"(pix)        // %3
   :
-  : "cc", "memory", "q0", "q1", "q2", "q3",
-    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
   );
 }
 #endif  // HAS_ARGBTOUV422ROW_NEON
@@ -1732,128 +1566,108 @@ void ARGBToUV422Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
 void ARGBToUV411Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
                          int pix) {
   asm volatile (
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-    ".p2align   2                              \n"
+    RGBTOUV_SETUP_REG
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
-    MEMACCESS(0)
-    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
-    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
     MEMACCESS(0)
-    "vld4.8     {d8, d10, d12, d14}, [%0]!     \n"  // load 8 more ARGB pixels.
-    MEMACCESS(0)
-    "vld4.8     {d9, d11, d13, d15}, [%0]!     \n"  // load last 8 ARGB pixels.
-    "vpaddl.u8  q4, q4                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q5, q5                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q6, q6                         \n"  // R 16 bytes -> 8 shorts.
+    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%0], #64 \n"  // load next 16.
+    "uaddlp     v4.8h, v4.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v5.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v6.8h, v6.16b                  \n"  // R 16 bytes -> 8 shorts.
 
-    "vpadd.u16  d0, d0, d1                     \n"  // B 16 shorts -> 8 shorts.
-    "vpadd.u16  d1, d8, d9                     \n"  // B
-    "vpadd.u16  d2, d2, d3                     \n"  // G 16 shorts -> 8 shorts.
-    "vpadd.u16  d3, d10, d11                   \n"  // G
-    "vpadd.u16  d4, d4, d5                     \n"  // R 16 shorts -> 8 shorts.
-    "vpadd.u16  d5, d12, d13                   \n"  // R
+    "addp       v0.8h, v0.8h, v4.8h            \n"  // B 16 shorts -> 8 shorts.
+    "addp       v1.8h, v1.8h, v5.8h            \n"  // G 16 shorts -> 8 shorts.
+    "addp       v2.8h, v2.8h, v6.8h            \n"  // R 16 shorts -> 8 shorts.
 
-    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
-    "vrshr.u16  q1, q1, #1                     \n"
-    "vrshr.u16  q2, q2, #1                     \n"
+    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
+    "urshr      v1.8h, v1.8h, #1               \n"
+    "urshr      v2.8h, v2.8h, #1               \n"
 
     "subs       %3, %3, #32                    \n"  // 32 processed per loop.
-    "vmul.s16   q8, q0, q10                    \n"  // B
-    "vmls.s16   q8, q1, q11                    \n"  // G
-    "vmls.s16   q8, q2, q12                    \n"  // R
-    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
-    "vmul.s16   q9, q2, q10                    \n"  // R
-    "vmls.s16   q9, q1, q14                    \n"  // G
-    "vmls.s16   q9, q0, q13                    \n"  // B
-    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
-    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
-    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels U.
+    "mul        v3.8h, v0.8h, v20.8h           \n"  // B
+    "mls        v3.8h, v1.8h, v21.8h           \n"  // G
+    "mls        v3.8h, v2.8h, v22.8h           \n"  // R
+    "add        v3.8h, v3.8h, v25.8h           \n"  // +128 -> unsigned
+    "mul        v4.8h, v2.8h, v20.8h           \n"  // R
+    "mls        v4.8h, v1.8h, v24.8h           \n"  // G
+    "mls        v4.8h, v0.8h, v23.8h           \n"  // B
+    "add        v4.8h, v4.8h, v25.8h           \n"  // +128 -> unsigned
+    "uqshrn     v0.8b, v3.8h, #8               \n"  // 16 bit to 8 bit U
+    "uqshrn     v1.8b, v4.8h, #8               \n"  // 16 bit to 8 bit V
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels U.
     MEMACCESS(2)
-    "vst1.8     {d1}, [%2]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
+    "st1        {v1.8b}, [%2], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
   : "+r"(src_argb),  // %0
     "+r"(dst_u),     // %1
     "+r"(dst_v),     // %2
     "+r"(pix)        // %3
   :
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
-    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
   );
 }
 #endif  // HAS_ARGBTOUV411ROW_NEON
 
 // 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
 #define RGBTOUV(QB, QG, QR) \
-    "vmul.s16   q8, " #QB ", q10               \n"  /* B                    */ \
-    "vmls.s16   q8, " #QG ", q11               \n"  /* G                    */ \
-    "vmls.s16   q8, " #QR ", q12               \n"  /* R                    */ \
-    "vadd.u16   q8, q8, q15                    \n"  /* +128 -> unsigned     */ \
-    "vmul.s16   q9, " #QR ", q10               \n"  /* R                    */ \
-    "vmls.s16   q9, " #QG ", q14               \n"  /* G                    */ \
-    "vmls.s16   q9, " #QB ", q13               \n"  /* B                    */ \
-    "vadd.u16   q9, q9, q15                    \n"  /* +128 -> unsigned     */ \
-    "vqshrn.u16  d0, q8, #8                    \n"  /* 16 bit to 8 bit U    */ \
-    "vqshrn.u16  d1, q9, #8                    \n"  /* 16 bit to 8 bit V    */
+    "mul        v3.8h, " #QB ",v20.8h          \n"  /* B                    */ \
+    "mul        v4.8h, " #QR ",v20.8h          \n"  /* R                    */ \
+    "mls        v3.8h, " #QG ",v21.8h          \n"  /* G                    */ \
+    "mls        v4.8h, " #QG ",v24.8h          \n"  /* G                    */ \
+    "mls        v3.8h, " #QR ",v22.8h          \n"  /* R                    */ \
+    "mls        v4.8h, " #QB ",v23.8h          \n"  /* B                    */ \
+    "add        v3.8h, v3.8h, v25.8h           \n"  /* +128 -> unsigned     */ \
+    "add        v4.8h, v4.8h, v25.8h           \n"  /* +128 -> unsigned     */ \
+    "uqshrn     v0.8b, v3.8h, #8               \n"  /* 16 bit to 8 bit U    */ \
+    "uqshrn     v1.8b, v4.8h, #8               \n"  /* 16 bit to 8 bit V    */
 
 // TODO(fbarchard): Consider vhadd vertical, then vpaddl horizontal, avoid shr.
+// TODO(fbarchard): consider ptrdiff_t for all strides.
+
 #ifdef HAS_ARGBTOUVROW_NEON
 void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb,
                       uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_argb_1 = src_argb + src_stride_argb;
   asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_argb
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-    ".p2align   2                              \n"
+    RGBTOUV_SETUP_REG
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
-    MEMACCESS(0)
-    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
-    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
-    MEMACCESS(1)
-    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ARGB pixels.
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
+
     MEMACCESS(1)
-    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ARGB pixels.
-    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
-    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
-    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
+    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load next 16
+    "uadalp     v0.8h, v4.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uadalp     v1.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uadalp     v2.8h, v6.16b                  \n"  // R 16 bytes -> 8 shorts.
 
-    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
-    "vrshr.u16  q1, q1, #1                     \n"
-    "vrshr.u16  q2, q2, #1                     \n"
+    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
+    "urshr      v1.8h, v1.8h, #1               \n"
+    "urshr      v2.8h, v2.8h, #1               \n"
 
     "subs       %4, %4, #16                    \n"  // 32 processed per loop.
-    RGBTOUV(q0, q1, q2)
+    RGBTOUV(v0.8h, v1.8h, v2.8h)
     MEMACCESS(2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
     MEMACCESS(3)
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
   : "+r"(src_argb),  // %0
-    "+r"(src_stride_argb),  // %1
+    "+r"(src_argb_1),  // %1
     "+r"(dst_u),     // %2
     "+r"(dst_v),     // %3
     "+r"(pix)        // %4
   :
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
-    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
   );
 }
 #endif  // HAS_ARGBTOUVROW_NEON
@@ -1862,50 +1676,45 @@ void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb,
 #ifdef HAS_ARGBTOUVJROW_NEON
 void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb,
                        uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_argb_1 = src_argb + src_stride_argb;
   asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_argb
-    "vmov.s16   q10, #127 / 2                  \n"  // UB / VR 0.500 coefficient
-    "vmov.s16   q11, #84 / 2                   \n"  // UG -0.33126 coefficient
-    "vmov.s16   q12, #43 / 2                   \n"  // UR -0.16874 coefficient
-    "vmov.s16   q13, #20 / 2                   \n"  // VB -0.08131 coefficient
-    "vmov.s16   q14, #107 / 2                  \n"  // VG -0.41869 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-    ".p2align   2                              \n"
+    "movi       v20.8h, #63, lsl #0            \n"  // UB/VR coeff (0.500) / 2
+    "movi       v21.8h, #42, lsl #0            \n"  // UG coeff (-0.33126) / 2
+    "movi       v22.8h, #21, lsl #0            \n"  // UR coeff (-0.16874) / 2
+    "movi       v23.8h, #10, lsl #0            \n"  // VB coeff (-0.08131) / 2
+    "movi       v24.8h, #53, lsl #0            \n"  // VG coeff (-0.41869) / 2
+    "movi       v25.16b, #0x80                 \n"  // 128.5 (0x8080 in 16-bit)
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
-    MEMACCESS(0)
-    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
-    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
     MEMACCESS(1)
-    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ARGB pixels.
-    MEMACCESS(1)
-    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ARGB pixels.
-    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
-    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
-    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
+    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64  \n"  // load next 16
+    "uadalp     v0.8h, v4.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uadalp     v1.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uadalp     v2.8h, v6.16b                  \n"  // R 16 bytes -> 8 shorts.
 
-    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
-    "vrshr.u16  q1, q1, #1                     \n"
-    "vrshr.u16  q2, q2, #1                     \n"
+    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
+    "urshr      v1.8h, v1.8h, #1               \n"
+    "urshr      v2.8h, v2.8h, #1               \n"
 
     "subs       %4, %4, #16                    \n"  // 32 processed per loop.
-    RGBTOUV(q0, q1, q2)
+    RGBTOUV(v0.8h, v1.8h, v2.8h)
     MEMACCESS(2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
     MEMACCESS(3)
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
   : "+r"(src_argb),  // %0
-    "+r"(src_stride_argb),  // %1
+    "+r"(src_argb_1),  // %1
     "+r"(dst_u),     // %2
     "+r"(dst_v),     // %3
     "+r"(pix)        // %4
   :
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
-    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
   );
 }
 #endif  // HAS_ARGBTOUVJROW_NEON
@@ -1913,50 +1722,40 @@ void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb,
 #ifdef HAS_BGRATOUVROW_NEON
 void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra,
                       uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_bgra_1 = src_bgra + src_stride_bgra;
   asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_bgra
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-    ".p2align   2                              \n"
+    RGBTOUV_SETUP_REG
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 BGRA pixels.
-    MEMACCESS(0)
-    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 BGRA pixels.
-    "vpaddl.u8  q3, q3                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q2, q2                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q1                         \n"  // R 16 bytes -> 8 shorts.
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+    "uaddlp     v0.8h, v3.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v3.8h, v2.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v1.16b                  \n"  // R 16 bytes -> 8 shorts.
     MEMACCESS(1)
-    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more BGRA pixels.
-    MEMACCESS(1)
-    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 BGRA pixels.
-    "vpadal.u8  q3, q7                         \n"  // B 16 bytes -> 8 shorts.
-    "vpadal.u8  q2, q6                         \n"  // G 16 bytes -> 8 shorts.
-    "vpadal.u8  q1, q5                         \n"  // R 16 bytes -> 8 shorts.
+    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load 16 more
+    "uadalp     v0.8h, v7.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uadalp     v3.8h, v6.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uadalp     v2.8h, v5.16b                  \n"  // R 16 bytes -> 8 shorts.
 
-    "vrshr.u16  q1, q1, #1                     \n"  // 2x average
-    "vrshr.u16  q2, q2, #1                     \n"
-    "vrshr.u16  q3, q3, #1                     \n"
+    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
+    "urshr      v1.8h, v3.8h, #1               \n"
+    "urshr      v2.8h, v2.8h, #1               \n"
 
     "subs       %4, %4, #16                    \n"  // 32 processed per loop.
-    RGBTOUV(q3, q2, q1)
+    RGBTOUV(v0.8h, v1.8h, v2.8h)
     MEMACCESS(2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
     MEMACCESS(3)
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
   : "+r"(src_bgra),  // %0
-    "+r"(src_stride_bgra),  // %1
+    "+r"(src_bgra_1),  // %1
     "+r"(dst_u),     // %2
     "+r"(dst_v),     // %3
     "+r"(pix)        // %4
   :
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
-    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
   );
 }
 #endif  // HAS_BGRATOUVROW_NEON
@@ -1964,50 +1763,40 @@ void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra,
 #ifdef HAS_ABGRTOUVROW_NEON
 void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr,
                       uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_abgr_1 = src_abgr + src_stride_abgr;
   asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_abgr
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-    ".p2align   2                              \n"
+    RGBTOUV_SETUP_REG
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ABGR pixels.
-    MEMACCESS(0)
-    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ABGR pixels.
-    "vpaddl.u8  q2, q2                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q0, q0                         \n"  // R 16 bytes -> 8 shorts.
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+    "uaddlp     v3.8h, v2.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v0.16b                  \n"  // R 16 bytes -> 8 shorts.
     MEMACCESS(1)
-    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ABGR pixels.
-    MEMACCESS(1)
-    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ABGR pixels.
-    "vpadal.u8  q2, q6                         \n"  // B 16 bytes -> 8 shorts.
-    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
-    "vpadal.u8  q0, q4                         \n"  // R 16 bytes -> 8 shorts.
+    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load 16 more.
+    "uadalp     v3.8h, v6.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uadalp     v2.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uadalp     v1.8h, v4.16b                  \n"  // R 16 bytes -> 8 shorts.
 
-    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
-    "vrshr.u16  q1, q1, #1                     \n"
-    "vrshr.u16  q2, q2, #1                     \n"
+    "urshr      v0.8h, v3.8h, #1               \n"  // 2x average
+    "urshr      v2.8h, v2.8h, #1               \n"
+    "urshr      v1.8h, v1.8h, #1               \n"
 
     "subs       %4, %4, #16                    \n"  // 32 processed per loop.
-    RGBTOUV(q2, q1, q0)
+    RGBTOUV(v0.8h, v2.8h, v1.8h)
     MEMACCESS(2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
     MEMACCESS(3)
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
   : "+r"(src_abgr),  // %0
-    "+r"(src_stride_abgr),  // %1
+    "+r"(src_abgr_1),  // %1
     "+r"(dst_u),     // %2
     "+r"(dst_v),     // %3
     "+r"(pix)        // %4
   :
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
-    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
   );
 }
 #endif  // HAS_ABGRTOUVROW_NEON
@@ -2015,50 +1804,40 @@ void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr,
 #ifdef HAS_RGBATOUVROW_NEON
 void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba,
                       uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_rgba_1 = src_rgba + src_stride_rgba;
   asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_rgba
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-    ".p2align   2                              \n"
+    RGBTOUV_SETUP_REG
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 RGBA pixels.
-    MEMACCESS(0)
-    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 RGBA pixels.
-    "vpaddl.u8  q0, q1                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q2                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q2, q3                         \n"  // R 16 bytes -> 8 shorts.
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+    "uaddlp     v0.8h, v1.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v2.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v3.16b                  \n"  // R 16 bytes -> 8 shorts.
     MEMACCESS(1)
-    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more RGBA pixels.
-    MEMACCESS(1)
-    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 RGBA pixels.
-    "vpadal.u8  q0, q5                         \n"  // B 16 bytes -> 8 shorts.
-    "vpadal.u8  q1, q6                         \n"  // G 16 bytes -> 8 shorts.
-    "vpadal.u8  q2, q7                         \n"  // R 16 bytes -> 8 shorts.
+    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load 16 more.
+    "uadalp     v0.8h, v5.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uadalp     v1.8h, v6.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uadalp     v2.8h, v7.16b                  \n"  // R 16 bytes -> 8 shorts.
 
-    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
-    "vrshr.u16  q1, q1, #1                     \n"
-    "vrshr.u16  q2, q2, #1                     \n"
+    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
+    "urshr      v1.8h, v1.8h, #1               \n"
+    "urshr      v2.8h, v2.8h, #1               \n"
 
     "subs       %4, %4, #16                    \n"  // 32 processed per loop.
-    RGBTOUV(q0, q1, q2)
+    RGBTOUV(v0.8h, v1.8h, v2.8h)
     MEMACCESS(2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
     MEMACCESS(3)
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
   : "+r"(src_rgba),  // %0
-    "+r"(src_stride_rgba),  // %1
+    "+r"(src_rgba_1),  // %1
     "+r"(dst_u),     // %2
     "+r"(dst_v),     // %3
     "+r"(pix)        // %4
   :
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
-    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
   );
 }
 #endif  // HAS_RGBATOUVROW_NEON
@@ -2066,50 +1845,40 @@ void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba,
 #ifdef HAS_RGB24TOUVROW_NEON
 void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24,
                        uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_rgb24_1 = src_rgb24 + src_stride_rgb24;
   asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_rgb24
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-    ".p2align   2                              \n"
+    RGBTOUV_SETUP_REG
   "1:                                          \n"
     MEMACCESS(0)
-    "vld3.8     {d0, d2, d4}, [%0]!            \n"  // load 8 RGB24 pixels.
-    MEMACCESS(0)
-    "vld3.8     {d1, d3, d5}, [%0]!            \n"  // load next 8 RGB24 pixels.
-    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
-    MEMACCESS(1)
-    "vld3.8     {d8, d10, d12}, [%1]!          \n"  // load 8 more RGB24 pixels.
+    "ld3        {v0.16b,v1.16b,v2.16b}, [%0], #48 \n"  // load 16 pixels.
+    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
     MEMACCESS(1)
-    "vld3.8     {d9, d11, d13}, [%1]!          \n"  // load last 8 RGB24 pixels.
-    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
-    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
-    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
+    "ld3        {v4.16b,v5.16b,v6.16b}, [%1], #48 \n"  // load 16 more.
+    "uadalp     v0.8h, v4.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uadalp     v1.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uadalp     v2.8h, v6.16b                  \n"  // R 16 bytes -> 8 shorts.
 
-    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
-    "vrshr.u16  q1, q1, #1                     \n"
-    "vrshr.u16  q2, q2, #1                     \n"
+    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
+    "urshr      v1.8h, v1.8h, #1               \n"
+    "urshr      v2.8h, v2.8h, #1               \n"
 
     "subs       %4, %4, #16                    \n"  // 32 processed per loop.
-    RGBTOUV(q0, q1, q2)
+    RGBTOUV(v0.8h, v1.8h, v2.8h)
     MEMACCESS(2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
     MEMACCESS(3)
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
   : "+r"(src_rgb24),  // %0
-    "+r"(src_stride_rgb24),  // %1
+    "+r"(src_rgb24_1),  // %1
     "+r"(dst_u),     // %2
     "+r"(dst_v),     // %3
     "+r"(pix)        // %4
   :
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
-    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
   );
 }
 #endif  // HAS_RGB24TOUVROW_NEON
@@ -2117,50 +1886,40 @@ void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24,
 #ifdef HAS_RAWTOUVROW_NEON
 void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw,
                      uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_raw_1 = src_raw + src_stride_raw;
   asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_raw
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-    ".p2align   2                              \n"
+    RGBTOUV_SETUP_REG
   "1:                                          \n"
     MEMACCESS(0)
-    "vld3.8     {d0, d2, d4}, [%0]!            \n"  // load 8 RAW pixels.
-    MEMACCESS(0)
-    "vld3.8     {d1, d3, d5}, [%0]!            \n"  // load next 8 RAW pixels.
-    "vpaddl.u8  q2, q2                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q0, q0                         \n"  // R 16 bytes -> 8 shorts.
-    MEMACCESS(1)
-    "vld3.8     {d8, d10, d12}, [%1]!          \n"  // load 8 more RAW pixels.
+    "ld3        {v0.16b,v1.16b,v2.16b}, [%0], #48 \n"  // load 8 RAW pixels.
+    "uaddlp     v2.8h, v2.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v0.8h, v0.16b                  \n"  // R 16 bytes -> 8 shorts.
     MEMACCESS(1)
-    "vld3.8     {d9, d11, d13}, [%1]!          \n"  // load last 8 RAW pixels.
-    "vpadal.u8  q2, q6                         \n"  // B 16 bytes -> 8 shorts.
-    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
-    "vpadal.u8  q0, q4                         \n"  // R 16 bytes -> 8 shorts.
+    "ld3        {v4.16b,v5.16b,v6.16b}, [%1], #48 \n"  // load 8 more RAW pixels
+    "uadalp     v2.8h, v6.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uadalp     v1.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uadalp     v0.8h, v4.16b                  \n"  // R 16 bytes -> 8 shorts.
 
-    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
-    "vrshr.u16  q1, q1, #1                     \n"
-    "vrshr.u16  q2, q2, #1                     \n"
+    "urshr      v2.8h, v2.8h, #1               \n"  // 2x average
+    "urshr      v1.8h, v1.8h, #1               \n"
+    "urshr      v0.8h, v0.8h, #1               \n"
 
     "subs       %4, %4, #16                    \n"  // 32 processed per loop.
-    RGBTOUV(q2, q1, q0)
+    RGBTOUV(v2.8h, v1.8h, v0.8h)
     MEMACCESS(2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
     MEMACCESS(3)
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
   : "+r"(src_raw),  // %0
-    "+r"(src_stride_raw),  // %1
+    "+r"(src_raw_1),  // %1
     "+r"(dst_u),     // %2
     "+r"(dst_v),     // %3
     "+r"(pix)        // %4
   :
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
-    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
   );
 }
 #endif  // HAS_RAWTOUVROW_NEON
@@ -2169,70 +1928,74 @@ void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw,
 #ifdef HAS_RGB565TOUVROW_NEON
 void RGB565ToUVRow_NEON(const uint8* src_rgb565, int src_stride_rgb565,
                         uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_rgb565_1 = src_rgb565 + src_stride_rgb565;
   asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_argb
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-    ".p2align   2                              \n"
+    "movi       v22.8h, #56, lsl #0            \n"  // UB / VR coeff (0.875) / 2
+    "movi       v23.8h, #37, lsl #0            \n"  // UG coeff (-0.5781) / 2
+    "movi       v24.8h, #19, lsl #0            \n"  // UR coeff (-0.2969) / 2
+    "movi       v25.8h, #9 , lsl #0            \n"  // VB coeff (-0.1406) / 2
+    "movi       v26.8h, #47, lsl #0            \n"  // VG coeff (-0.7344) / 2
+    "movi       v27.16b, #0x80                 \n"  // 128.5 (0x8080 in 16-bit)
   "1:                                          \n"
     MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 RGB565 pixels.
     RGB565TOARGB
-    "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
-    "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
-    "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    "uaddlp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uaddlp     v18.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uaddlp     v20.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
     MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // next 8 RGB565 pixels.
+    "ld1        {v0.16b}, [%0], #16            \n"  // next 8 RGB565 pixels.
     RGB565TOARGB
-    "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
-    "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
-    "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+    "uaddlp     v17.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uaddlp     v19.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uaddlp     v21.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
 
     MEMACCESS(1)
-    "vld1.8     {q0}, [%1]!                    \n"  // load 8 RGB565 pixels.
+    "ld1        {v0.16b}, [%1], #16            \n"  // load 8 RGB565 pixels.
     RGB565TOARGB
-    "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
-    "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
-    "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    "uadalp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uadalp     v18.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uadalp     v20.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
     MEMACCESS(1)
-    "vld1.8     {q0}, [%1]!                    \n"  // next 8 RGB565 pixels.
+    "ld1        {v0.16b}, [%1], #16            \n"  // next 8 RGB565 pixels.
     RGB565TOARGB
-    "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
-    "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
-    "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+    "uadalp     v17.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uadalp     v19.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uadalp     v21.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
 
-    "vrshr.u16  q4, q4, #1                     \n"  // 2x average
-    "vrshr.u16  q5, q5, #1                     \n"
-    "vrshr.u16  q6, q6, #1                     \n"
+    "ins        v16.D[1], v17.D[0]             \n"
+    "ins        v18.D[1], v19.D[0]             \n"
+    "ins        v20.D[1], v21.D[0]             \n"
+
+    "urshr      v4.8h, v16.8h, #1              \n"  // 2x average
+    "urshr      v5.8h, v18.8h, #1              \n"
+    "urshr      v6.8h, v20.8h, #1              \n"
 
     "subs       %4, %4, #16                    \n"  // 16 processed per loop.
-    "vmul.s16   q8, q4, q10                    \n"  // B
-    "vmls.s16   q8, q5, q11                    \n"  // G
-    "vmls.s16   q8, q6, q12                    \n"  // R
-    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
-    "vmul.s16   q9, q6, q10                    \n"  // R
-    "vmls.s16   q9, q5, q14                    \n"  // G
-    "vmls.s16   q9, q4, q13                    \n"  // B
-    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
-    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
-    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    "mul        v16.8h, v4.8h, v22.8h          \n"  // B
+    "mls        v16.8h, v5.8h, v23.8h          \n"  // G
+    "mls        v16.8h, v6.8h, v24.8h          \n"  // R
+    "add        v16.8h, v16.8h, v27.8h         \n"  // +128 -> unsigned
+    "mul        v17.8h, v6.8h, v22.8h          \n"  // R
+    "mls        v17.8h, v5.8h, v26.8h          \n"  // G
+    "mls        v17.8h, v4.8h, v25.8h          \n"  // B
+    "add        v17.8h, v17.8h, v27.8h         \n"  // +128 -> unsigned
+    "uqshrn     v0.8b, v16.8h, #8              \n"  // 16 bit to 8 bit U
+    "uqshrn     v1.8b, v17.8h, #8              \n"  // 16 bit to 8 bit V
     MEMACCESS(2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
     MEMACCESS(3)
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
   : "+r"(src_rgb565),  // %0
-    "+r"(src_stride_rgb565),  // %1
+    "+r"(src_rgb565_1),  // %1
     "+r"(dst_u),     // %2
     "+r"(dst_v),     // %3
     "+r"(pix)        // %4
   :
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
-    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24",
+    "v25", "v26", "v27"
   );
 }
 #endif  // HAS_RGB565TOUVROW_NEON
@@ -2241,70 +2004,69 @@ void RGB565ToUVRow_NEON(const uint8* src_rgb565, int src_stride_rgb565,
 #ifdef HAS_ARGB1555TOUVROW_NEON
 void ARGB1555ToUVRow_NEON(const uint8* src_argb1555, int src_stride_argb1555,
                         uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_argb1555_1 = src_argb1555 + src_stride_argb1555;
   asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_argb
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-    ".p2align   2                              \n"
+    RGBTOUV_SETUP_REG
   "1:                                          \n"
     MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB1555 pixels.
     RGB555TOARGB
-    "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
-    "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
-    "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    "uaddlp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uaddlp     v17.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uaddlp     v18.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
     MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // next 8 ARGB1555 pixels.
+    "ld1        {v0.16b}, [%0], #16            \n"  // next 8 ARGB1555 pixels.
     RGB555TOARGB
-    "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
-    "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
-    "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+    "uaddlp     v26.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uaddlp     v27.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uaddlp     v28.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
 
     MEMACCESS(1)
-    "vld1.8     {q0}, [%1]!                    \n"  // load 8 ARGB1555 pixels.
+    "ld1        {v0.16b}, [%1], #16            \n"  // load 8 ARGB1555 pixels.
     RGB555TOARGB
-    "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
-    "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
-    "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    "uadalp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uadalp     v17.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uadalp     v18.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
     MEMACCESS(1)
-    "vld1.8     {q0}, [%1]!                    \n"  // next 8 ARGB1555 pixels.
+    "ld1        {v0.16b}, [%1], #16            \n"  // next 8 ARGB1555 pixels.
     RGB555TOARGB
-    "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
-    "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
-    "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+    "uadalp     v26.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uadalp     v27.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uadalp     v28.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+
+    "ins        v16.D[1], v26.D[0]             \n"
+    "ins        v17.D[1], v27.D[0]             \n"
+    "ins        v18.D[1], v28.D[0]             \n"
 
-    "vrshr.u16  q4, q4, #1                     \n"  // 2x average
-    "vrshr.u16  q5, q5, #1                     \n"
-    "vrshr.u16  q6, q6, #1                     \n"
+    "urshr      v4.8h, v16.8h, #1              \n"  // 2x average
+    "urshr      v5.8h, v17.8h, #1              \n"
+    "urshr      v6.8h, v18.8h, #1              \n"
 
     "subs       %4, %4, #16                    \n"  // 16 processed per loop.
-    "vmul.s16   q8, q4, q10                    \n"  // B
-    "vmls.s16   q8, q5, q11                    \n"  // G
-    "vmls.s16   q8, q6, q12                    \n"  // R
-    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
-    "vmul.s16   q9, q6, q10                    \n"  // R
-    "vmls.s16   q9, q5, q14                    \n"  // G
-    "vmls.s16   q9, q4, q13                    \n"  // B
-    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
-    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
-    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    "mul        v2.8h, v4.8h, v20.8h           \n"  // B
+    "mls        v2.8h, v5.8h, v21.8h           \n"  // G
+    "mls        v2.8h, v6.8h, v22.8h           \n"  // R
+    "add        v2.8h, v2.8h, v25.8h           \n"  // +128 -> unsigned
+    "mul        v3.8h, v6.8h, v20.8h           \n"  // R
+    "mls        v3.8h, v5.8h, v24.8h           \n"  // G
+    "mls        v3.8h, v4.8h, v23.8h           \n"  // B
+    "add        v3.8h, v3.8h, v25.8h           \n"  // +128 -> unsigned
+    "uqshrn     v0.8b, v2.8h, #8               \n"  // 16 bit to 8 bit U
+    "uqshrn     v1.8b, v3.8h, #8               \n"  // 16 bit to 8 bit V
     MEMACCESS(2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
     MEMACCESS(3)
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
   : "+r"(src_argb1555),  // %0
-    "+r"(src_stride_argb1555),  // %1
+    "+r"(src_argb1555_1),  // %1
     "+r"(dst_u),     // %2
     "+r"(dst_v),     // %3
     "+r"(pix)        // %4
   :
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
-    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6",
+    "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25",
+    "v26", "v27", "v28"
   );
 }
 #endif  // HAS_ARGB1555TOUVROW_NEON
@@ -2313,70 +2075,70 @@ void ARGB1555ToUVRow_NEON(const uint8* src_argb1555, int src_stride_argb1555,
 #ifdef HAS_ARGB4444TOUVROW_NEON
 void ARGB4444ToUVRow_NEON(const uint8* src_argb4444, int src_stride_argb4444,
                           uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_argb4444_1 = src_argb4444 + src_stride_argb4444;
   asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_argb
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-    ".p2align   2                              \n"
+    RGBTOUV_SETUP_REG
   "1:                                          \n"
     MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB4444 pixels.
     ARGB4444TOARGB
-    "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
-    "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
-    "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    "uaddlp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uaddlp     v17.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uaddlp     v18.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
     MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // next 8 ARGB4444 pixels.
+    "ld1        {v0.16b}, [%0], #16            \n"  // next 8 ARGB4444 pixels.
     ARGB4444TOARGB
-    "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
-    "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
-    "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+    "uaddlp     v26.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uaddlp     v27.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uaddlp     v28.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
 
     MEMACCESS(1)
-    "vld1.8     {q0}, [%1]!                    \n"  // load 8 ARGB4444 pixels.
+    "ld1        {v0.16b}, [%1], #16            \n"  // load 8 ARGB4444 pixels.
     ARGB4444TOARGB
-    "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
-    "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
-    "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    "uadalp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uadalp     v17.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uadalp     v18.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
     MEMACCESS(1)
-    "vld1.8     {q0}, [%1]!                    \n"  // next 8 ARGB4444 pixels.
+    "ld1        {v0.16b}, [%1], #16            \n"  // next 8 ARGB4444 pixels.
     ARGB4444TOARGB
-    "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
-    "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
-    "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+    "uadalp     v26.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uadalp     v27.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uadalp     v28.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
 
-    "vrshr.u16  q4, q4, #1                     \n"  // 2x average
-    "vrshr.u16  q5, q5, #1                     \n"
-    "vrshr.u16  q6, q6, #1                     \n"
+    "ins        v16.D[1], v26.D[0]             \n"
+    "ins        v17.D[1], v27.D[0]             \n"
+    "ins        v18.D[1], v28.D[0]             \n"
+
+    "urshr      v4.8h, v16.8h, #1              \n"  // 2x average
+    "urshr      v5.8h, v17.8h, #1              \n"
+    "urshr      v6.8h, v18.8h, #1              \n"
 
     "subs       %4, %4, #16                    \n"  // 16 processed per loop.
-    "vmul.s16   q8, q4, q10                    \n"  // B
-    "vmls.s16   q8, q5, q11                    \n"  // G
-    "vmls.s16   q8, q6, q12                    \n"  // R
-    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
-    "vmul.s16   q9, q6, q10                    \n"  // R
-    "vmls.s16   q9, q5, q14                    \n"  // G
-    "vmls.s16   q9, q4, q13                    \n"  // B
-    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
-    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
-    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    "mul        v2.8h, v4.8h, v20.8h           \n"  // B
+    "mls        v2.8h, v5.8h, v21.8h           \n"  // G
+    "mls        v2.8h, v6.8h, v22.8h           \n"  // R
+    "add        v2.8h, v2.8h, v25.8h           \n"  // +128 -> unsigned
+    "mul        v3.8h, v6.8h, v20.8h           \n"  // R
+    "mls        v3.8h, v5.8h, v24.8h           \n"  // G
+    "mls        v3.8h, v4.8h, v23.8h           \n"  // B
+    "add        v3.8h, v3.8h, v25.8h           \n"  // +128 -> unsigned
+    "uqshrn     v0.8b, v2.8h, #8               \n"  // 16 bit to 8 bit U
+    "uqshrn     v1.8b, v3.8h, #8               \n"  // 16 bit to 8 bit V
     MEMACCESS(2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
     MEMACCESS(3)
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
   : "+r"(src_argb4444),  // %0
-    "+r"(src_stride_argb4444),  // %1
+    "+r"(src_argb4444_1),  // %1
     "+r"(dst_u),     // %2
     "+r"(dst_v),     // %3
     "+r"(pix)        // %4
   :
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
-    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6",
+    "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25",
+    "v26", "v27", "v28"
+
   );
 }
 #endif  // HAS_ARGB4444TOUVROW_NEON
@@ -2384,29 +2146,29 @@ void ARGB4444ToUVRow_NEON(const uint8* src_argb4444, int src_stride_argb4444,
 #ifdef HAS_RGB565TOYROW_NEON
 void RGB565ToYRow_NEON(const uint8* src_rgb565, uint8* dst_y, int pix) {
   asm volatile (
-    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
-    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
-    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
-    "vmov.u8    d27, #16                       \n"  // Add 16 constant
-    ".p2align   2                              \n"
+    "movi       v24.8b, #13                    \n"  // B * 0.1016 coefficient
+    "movi       v25.8b, #65                    \n"  // G * 0.5078 coefficient
+    "movi       v26.8b, #33                    \n"  // R * 0.2578 coefficient
+    "movi       v27.8b, #16                    \n"  // Add 16 constant
   "1:                                          \n"
     MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 RGB565 pixels.
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
     RGB565TOARGB
-    "vmull.u8   q2, d0, d24                    \n"  // B
-    "vmlal.u8   q2, d1, d25                    \n"  // G
-    "vmlal.u8   q2, d2, d26                    \n"  // R
-    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
-    "vqadd.u8   d0, d27                        \n"
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-    "bgt        1b                             \n"
+    "umull      v3.8h, v0.8b, v24.8b           \n"  // B
+    "umlal      v3.8h, v1.8b, v25.8b           \n"  // G
+    "umlal      v3.8h, v2.8b, v26.8b           \n"  // R
+    "sqrshrun   v0.8b, v3.8h, #7               \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v27.8b           \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
   : "+r"(src_rgb565),  // %0
     "+r"(dst_y),       // %1
     "+r"(pix)          // %2
   :
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v6",
+    "v24", "v25", "v26", "v27"
   );
 }
 #endif  // HAS_RGB565TOYROW_NEON
@@ -2414,29 +2176,28 @@ void RGB565ToYRow_NEON(const uint8* src_rgb565, uint8* dst_y, int pix) {
 #ifdef HAS_ARGB1555TOYROW_NEON
 void ARGB1555ToYRow_NEON(const uint8* src_argb1555, uint8* dst_y, int pix) {
   asm volatile (
-    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
-    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
-    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
-    "vmov.u8    d27, #16                       \n"  // Add 16 constant
-    ".p2align   2                              \n"
+    "movi       v4.8b, #13                     \n"  // B * 0.1016 coefficient
+    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
+    "movi       v6.8b, #33                     \n"  // R * 0.2578 coefficient
+    "movi       v7.8b, #16                     \n"  // Add 16 constant
   "1:                                          \n"
     MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB1555 pixels.
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
     ARGB1555TOARGB
-    "vmull.u8   q2, d0, d24                    \n"  // B
-    "vmlal.u8   q2, d1, d25                    \n"  // G
-    "vmlal.u8   q2, d2, d26                    \n"  // R
-    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
-    "vqadd.u8   d0, d27                        \n"
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-    "bgt        1b                             \n"
+    "umull      v3.8h, v0.8b, v4.8b            \n"  // B
+    "umlal      v3.8h, v1.8b, v5.8b            \n"  // G
+    "umlal      v3.8h, v2.8b, v6.8b            \n"  // R
+    "sqrshrun   v0.8b, v3.8h, #7               \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v7.8b            \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
   : "+r"(src_argb1555),  // %0
     "+r"(dst_y),         // %1
     "+r"(pix)            // %2
   :
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
   );
 }
 #endif  // HAS_ARGB1555TOYROW_NEON
@@ -2444,29 +2205,28 @@ void ARGB1555ToYRow_NEON(const uint8* src_argb1555, uint8* dst_y, int pix) {
 #ifdef HAS_ARGB4444TOYROW_NEON
 void ARGB4444ToYRow_NEON(const uint8* src_argb4444, uint8* dst_y, int pix) {
   asm volatile (
-    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
-    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
-    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
-    "vmov.u8    d27, #16                       \n"  // Add 16 constant
-    ".p2align   2                              \n"
+    "movi       v24.8b, #13                    \n"  // B * 0.1016 coefficient
+    "movi       v25.8b, #65                    \n"  // G * 0.5078 coefficient
+    "movi       v26.8b, #33                    \n"  // R * 0.2578 coefficient
+    "movi       v27.8b, #16                    \n"  // Add 16 constant
   "1:                                          \n"
     MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB4444 pixels.
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
     ARGB4444TOARGB
-    "vmull.u8   q2, d0, d24                    \n"  // B
-    "vmlal.u8   q2, d1, d25                    \n"  // G
-    "vmlal.u8   q2, d2, d26                    \n"  // R
-    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
-    "vqadd.u8   d0, d27                        \n"
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-    "bgt        1b                             \n"
+    "umull      v3.8h, v0.8b, v24.8b           \n"  // B
+    "umlal      v3.8h, v1.8b, v25.8b           \n"  // G
+    "umlal      v3.8h, v2.8b, v26.8b           \n"  // R
+    "sqrshrun   v0.8b, v3.8h, #7               \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v27.8b           \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
   : "+r"(src_argb4444),  // %0
     "+r"(dst_y),         // %1
     "+r"(pix)            // %2
   :
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v24", "v25", "v26", "v27"
   );
 }
 #endif  // HAS_ARGB4444TOYROW_NEON
@@ -2474,28 +2234,27 @@ void ARGB4444ToYRow_NEON(const uint8* src_argb4444, uint8* dst_y, int pix) {
 #ifdef HAS_BGRATOYROW_NEON
 void BGRAToYRow_NEON(const uint8* src_bgra, uint8* dst_y, int pix) {
   asm volatile (
-    "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
-    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
-    "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
-    "vmov.u8    d7, #16                        \n"  // Add 16 constant
-    ".p2align   2                              \n"
+    "movi       v4.8b, #33                     \n"  // R * 0.2578 coefficient
+    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
+    "movi       v6.8b, #13                     \n"  // B * 0.1016 coefficient
+    "movi       v7.8b, #16                     \n"  // Add 16 constant
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of BGRA.
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 pixels.
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q8, d1, d4                     \n"  // R
-    "vmlal.u8   q8, d2, d5                     \n"  // G
-    "vmlal.u8   q8, d3, d6                     \n"  // B
-    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
-    "vqadd.u8   d0, d7                         \n"
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-    "bgt        1b                             \n"
+    "umull      v16.8h, v1.8b, v4.8b           \n"  // R
+    "umlal      v16.8h, v2.8b, v5.8b           \n"  // G
+    "umlal      v16.8h, v3.8b, v6.8b           \n"  // B
+    "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v7.8b            \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
   : "+r"(src_bgra),  // %0
     "+r"(dst_y),     // %1
     "+r"(pix)        // %2
   :
-  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"
   );
 }
 #endif  // HAS_BGRATOYROW_NEON
@@ -2503,28 +2262,27 @@ void BGRAToYRow_NEON(const uint8* src_bgra, uint8* dst_y, int pix) {
 #ifdef HAS_ABGRTOYROW_NEON
 void ABGRToYRow_NEON(const uint8* src_abgr, uint8* dst_y, int pix) {
   asm volatile (
-    "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
-    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
-    "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
-    "vmov.u8    d7, #16                        \n"  // Add 16 constant
-    ".p2align   2                              \n"
+    "movi       v4.8b, #33                     \n"  // R * 0.2578 coefficient
+    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
+    "movi       v6.8b, #13                     \n"  // B * 0.1016 coefficient
+    "movi       v7.8b, #16                     \n"  // Add 16 constant
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ABGR.
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 pixels.
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q8, d0, d4                     \n"  // R
-    "vmlal.u8   q8, d1, d5                     \n"  // G
-    "vmlal.u8   q8, d2, d6                     \n"  // B
-    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
-    "vqadd.u8   d0, d7                         \n"
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-    "bgt        1b                             \n"
+    "umull      v16.8h, v0.8b, v4.8b           \n"  // R
+    "umlal      v16.8h, v1.8b, v5.8b           \n"  // G
+    "umlal      v16.8h, v2.8b, v6.8b           \n"  // B
+    "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v7.8b            \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
   : "+r"(src_abgr),  // %0
-    "+r"(dst_y),  // %1
+    "+r"(dst_y),     // %1
     "+r"(pix)        // %2
   :
-  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"
   );
 }
 #endif  // HAS_ABGRTOYROW_NEON
@@ -2532,28 +2290,27 @@ void ABGRToYRow_NEON(const uint8* src_abgr, uint8* dst_y, int pix) {
 #ifdef HAS_RGBATOYROW_NEON
 void RGBAToYRow_NEON(const uint8* src_rgba, uint8* dst_y, int pix) {
   asm volatile (
-    "vmov.u8    d4, #13                        \n"  // B * 0.1016 coefficient
-    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
-    "vmov.u8    d6, #33                        \n"  // R * 0.2578 coefficient
-    "vmov.u8    d7, #16                        \n"  // Add 16 constant
-    ".p2align   2                              \n"
+    "movi       v4.8b, #13                     \n"  // B * 0.1016 coefficient
+    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
+    "movi       v6.8b, #33                     \n"  // R * 0.2578 coefficient
+    "movi       v7.8b, #16                     \n"  // Add 16 constant
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of RGBA.
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 pixels.
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q8, d1, d4                     \n"  // B
-    "vmlal.u8   q8, d2, d5                     \n"  // G
-    "vmlal.u8   q8, d3, d6                     \n"  // R
-    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
-    "vqadd.u8   d0, d7                         \n"
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-    "bgt        1b                             \n"
+    "umull      v16.8h, v1.8b, v4.8b           \n"  // B
+    "umlal      v16.8h, v2.8b, v5.8b           \n"  // G
+    "umlal      v16.8h, v3.8b, v6.8b           \n"  // R
+    "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v7.8b            \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
   : "+r"(src_rgba),  // %0
-    "+r"(dst_y),  // %1
+    "+r"(dst_y),     // %1
     "+r"(pix)        // %2
   :
-  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"
   );
 }
 #endif  // HAS_RGBATOYROW_NEON
@@ -2561,28 +2318,27 @@ void RGBAToYRow_NEON(const uint8* src_rgba, uint8* dst_y, int pix) {
 #ifdef HAS_RGB24TOYROW_NEON
 void RGB24ToYRow_NEON(const uint8* src_rgb24, uint8* dst_y, int pix) {
   asm volatile (
-    "vmov.u8    d4, #13                        \n"  // B * 0.1016 coefficient
-    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
-    "vmov.u8    d6, #33                        \n"  // R * 0.2578 coefficient
-    "vmov.u8    d7, #16                        \n"  // Add 16 constant
-    ".p2align   2                              \n"
+    "movi       v4.8b, #13                     \n"  // B * 0.1016 coefficient
+    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
+    "movi       v6.8b, #33                     \n"  // R * 0.2578 coefficient
+    "movi       v7.8b, #16                     \n"  // Add 16 constant
   "1:                                          \n"
     MEMACCESS(0)
-    "vld3.8     {d0, d1, d2}, [%0]!            \n"  // load 8 pixels of RGB24.
+    "ld3        {v0.8b,v1.8b,v2.8b}, [%0], #24 \n"  // load 8 pixels.
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q8, d0, d4                     \n"  // B
-    "vmlal.u8   q8, d1, d5                     \n"  // G
-    "vmlal.u8   q8, d2, d6                     \n"  // R
-    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
-    "vqadd.u8   d0, d7                         \n"
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-    "bgt        1b                             \n"
+    "umull      v16.8h, v0.8b, v4.8b           \n"  // B
+    "umlal      v16.8h, v1.8b, v5.8b           \n"  // G
+    "umlal      v16.8h, v2.8b, v6.8b           \n"  // R
+    "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v7.8b            \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
   : "+r"(src_rgb24),  // %0
-    "+r"(dst_y),  // %1
-    "+r"(pix)        // %2
+    "+r"(dst_y),      // %1
+    "+r"(pix)         // %2
   :
-  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"
   );
 }
 #endif  // HAS_RGB24TOYROW_NEON
@@ -2590,28 +2346,27 @@ void RGB24ToYRow_NEON(const uint8* src_rgb24, uint8* dst_y, int pix) {
 #ifdef HAS_RAWTOYROW_NEON
 void RAWToYRow_NEON(const uint8* src_raw, uint8* dst_y, int pix) {
   asm volatile (
-    "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
-    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
-    "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
-    "vmov.u8    d7, #16                        \n"  // Add 16 constant
-    ".p2align   2                              \n"
+    "movi       v4.8b, #33                     \n"  // R * 0.2578 coefficient
+    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
+    "movi       v6.8b, #13                     \n"  // B * 0.1016 coefficient
+    "movi       v7.8b, #16                     \n"  // Add 16 constant
   "1:                                          \n"
     MEMACCESS(0)
-    "vld3.8     {d0, d1, d2}, [%0]!            \n"  // load 8 pixels of RAW.
+    "ld3        {v0.8b,v1.8b,v2.8b}, [%0], #24 \n"  // load 8 pixels.
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q8, d0, d4                     \n"  // B
-    "vmlal.u8   q8, d1, d5                     \n"  // G
-    "vmlal.u8   q8, d2, d6                     \n"  // R
-    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
-    "vqadd.u8   d0, d7                         \n"
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-    "bgt        1b                             \n"
+    "umull      v16.8h, v0.8b, v4.8b           \n"  // B
+    "umlal      v16.8h, v1.8b, v5.8b           \n"  // G
+    "umlal      v16.8h, v2.8b, v6.8b           \n"  // R
+    "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v7.8b            \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
   : "+r"(src_raw),  // %0
-    "+r"(dst_y),  // %1
-    "+r"(pix)        // %2
+    "+r"(dst_y),    // %1
+    "+r"(pix)       // %2
   :
-  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"
   );
 }
 #endif  // HAS_RAWTOYROW_NEON
@@ -2621,96 +2376,98 @@ void RAWToYRow_NEON(const uint8* src_raw, uint8* dst_y, int pix) {
 void InterpolateRow_NEON(uint8* dst_ptr,
                          const uint8* src_ptr, ptrdiff_t src_stride,
                          int dst_width, int source_y_fraction) {
+  int y1_fraction = source_y_fraction;
+  int y0_fraction = 256 - y1_fraction;
+  const uint8* src_ptr1 = src_ptr + src_stride;
   asm volatile (
     "cmp        %4, #0                         \n"
-    "beq        100f                           \n"
-    "add        %2, %1                         \n"
+    "b.eq       100f                           \n"
     "cmp        %4, #64                        \n"
-    "beq        75f                            \n"
+    "b.eq       75f                            \n"
     "cmp        %4, #128                       \n"
-    "beq        50f                            \n"
+    "b.eq       50f                            \n"
     "cmp        %4, #192                       \n"
-    "beq        25f                            \n"
+    "b.eq       25f                            \n"
 
-    "vdup.8     d5, %4                         \n"
-    "rsb        %4, #256                       \n"
-    "vdup.8     d4, %4                         \n"
+    "dup        v5.16b, %w4                    \n"
+    "dup        v4.16b, %w5                    \n"
     // General purpose row blend.
   "1:                                          \n"
     MEMACCESS(1)
-    "vld1.8     {q0}, [%1]!                    \n"
+    "ld1        {v0.16b}, [%1], #16            \n"
     MEMACCESS(2)
-    "vld1.8     {q1}, [%2]!                    \n"
+    "ld1        {v1.16b}, [%2], #16            \n"
     "subs       %3, %3, #16                    \n"
-    "vmull.u8   q13, d0, d4                    \n"
-    "vmull.u8   q14, d1, d4                    \n"
-    "vmlal.u8   q13, d2, d5                    \n"
-    "vmlal.u8   q14, d3, d5                    \n"
-    "vrshrn.u16 d0, q13, #8                    \n"
-    "vrshrn.u16 d1, q14, #8                    \n"
-    MEMACCESS(0)
-    "vst1.8     {q0}, [%0]!                    \n"
-    "bgt        1b                             \n"
+    "umull      v2.8h, v0.8b,  v4.8b           \n"
+    "umull2     v3.8h, v0.16b, v4.16b          \n"
+    "umlal      v2.8h, v1.8b,  v5.8b           \n"
+    "umlal2     v3.8h, v1.16b, v5.16b          \n"
+    "rshrn      v0.8b,  v2.8h, #8              \n"
+    "rshrn2     v0.16b, v3.8h, #8              \n"
+    MEMACCESS(0)
+    "st1        {v0.16b}, [%0], #16            \n"
+    "b.gt       1b                             \n"
     "b          99f                            \n"
 
     // Blend 25 / 75.
   "25:                                         \n"
     MEMACCESS(1)
-    "vld1.8     {q0}, [%1]!                    \n"
+    "ld1        {v0.16b}, [%1], #16            \n"
     MEMACCESS(2)
-    "vld1.8     {q1}, [%2]!                    \n"
+    "ld1        {v1.16b}, [%2], #16            \n"
     "subs       %3, %3, #16                    \n"
-    "vrhadd.u8  q0, q1                         \n"
-    "vrhadd.u8  q0, q1                         \n"
+    "urhadd     v0.16b, v0.16b, v1.16b         \n"
+    "urhadd     v0.16b, v0.16b, v1.16b         \n"
     MEMACCESS(0)
-    "vst1.8     {q0}, [%0]!                    \n"
-    "bgt        25b                            \n"
+    "st1        {v0.16b}, [%0], #16            \n"
+    "b.gt       25b                            \n"
     "b          99f                            \n"
 
     // Blend 50 / 50.
   "50:                                         \n"
     MEMACCESS(1)
-    "vld1.8     {q0}, [%1]!                    \n"
+    "ld1        {v0.16b}, [%1], #16            \n"
     MEMACCESS(2)
-    "vld1.8     {q1}, [%2]!                    \n"
+    "ld1        {v1.16b}, [%2], #16            \n"
     "subs       %3, %3, #16                    \n"
-    "vrhadd.u8  q0, q1                         \n"
+    "urhadd     v0.16b, v0.16b, v1.16b         \n"
     MEMACCESS(0)
-    "vst1.8     {q0}, [%0]!                    \n"
-    "bgt        50b                            \n"
+    "st1        {v0.16b}, [%0], #16            \n"
+    "b.gt       50b                            \n"
     "b          99f                            \n"
 
     // Blend 75 / 25.
   "75:                                         \n"
     MEMACCESS(1)
-    "vld1.8     {q1}, [%1]!                    \n"
+    "ld1        {v1.16b}, [%1], #16            \n"
     MEMACCESS(2)
-    "vld1.8     {q0}, [%2]!                    \n"
+    "ld1        {v0.16b}, [%2], #16            \n"
     "subs       %3, %3, #16                    \n"
-    "vrhadd.u8  q0, q1                         \n"
-    "vrhadd.u8  q0, q1                         \n"
+    "urhadd     v0.16b, v0.16b, v1.16b         \n"
+    "urhadd     v0.16b, v0.16b, v1.16b         \n"
     MEMACCESS(0)
-    "vst1.8     {q0}, [%0]!                    \n"
-    "bgt        75b                            \n"
+    "st1        {v0.16b}, [%0], #16            \n"
+    "b.gt       75b                            \n"
     "b          99f                            \n"
 
     // Blend 100 / 0 - Copy row unchanged.
   "100:                                        \n"
     MEMACCESS(1)
-    "vld1.8     {q0}, [%1]!                    \n"
+    "ld1        {v0.16b}, [%1], #16            \n"
     "subs       %3, %3, #16                    \n"
     MEMACCESS(0)
-    "vst1.8     {q0}, [%0]!                    \n"
-    "bgt        100b                           \n"
+    "st1        {v0.16b}, [%0], #16            \n"
+    "b.gt       100b                           \n"
 
   "99:                                         \n"
   : "+r"(dst_ptr),          // %0
     "+r"(src_ptr),          // %1
-    "+r"(src_stride),       // %2
+    "+r"(src_ptr1),         // %2
     "+r"(dst_width),        // %3
-    "+r"(source_y_fraction) // %4
+    "+r"(y1_fraction),      // %4
+    "+r"(y0_fraction)       // %5
   :
-  : "cc", "memory", "q0", "q1", "d4", "d5", "q13", "q14"
+  : "cc", "memory", "v0", "v1", "v3", "v4", "v5"
   );
 }
 #endif  // HAS_INTERPOLATEROW_NEON
@@ -2720,55 +2477,59 @@ void InterpolateRow_NEON(uint8* dst_ptr,
 void ARGBBlendRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
                        uint8* dst_argb, int width) {
   asm volatile (
-    "subs       %3, #8                         \n"
-    "blt        89f                            \n"
+    "subs       %3, %3, #8                     \n"
+    "b.lt       89f                            \n"
     // Blend 8 pixels.
   "8:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ARGB0.
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB0 pixels
     MEMACCESS(1)
-    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 pixels of ARGB1.
+    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 ARGB1 pixels
     "subs       %3, %3, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q10, d4, d3                    \n"  // db * a
-    "vmull.u8   q11, d5, d3                    \n"  // dg * a
-    "vmull.u8   q12, d6, d3                    \n"  // dr * a
-    "vqrshrn.u16 d20, q10, #8                  \n"  // db >>= 8
-    "vqrshrn.u16 d21, q11, #8                  \n"  // dg >>= 8
-    "vqrshrn.u16 d22, q12, #8                  \n"  // dr >>= 8
-    "vqsub.u8   q2, q2, q10                    \n"  // dbg - dbg * a / 256
-    "vqsub.u8   d6, d6, d22                    \n"  // dr - dr * a / 256
-    "vqadd.u8   q0, q0, q2                     \n"  // + sbg
-    "vqadd.u8   d2, d2, d6                     \n"  // + sr
-    "vmov.u8    d3, #255                       \n"  // a = 255
+    "umull      v16.8h, v4.8b, v3.8b           \n"  // db * a
+    "umull      v17.8h, v5.8b, v3.8b           \n"  // dg * a
+    "umull      v18.8h, v6.8b, v3.8b           \n"  // dr * a
+    "uqrshrn    v16.8b, v16.8h, #8             \n"  // db >>= 8
+    "uqrshrn    v17.8b, v17.8h, #8             \n"  // dg >>= 8
+    "uqrshrn    v18.8b, v18.8h, #8             \n"  // dr >>= 8
+    "uqsub      v4.8b, v4.8b, v16.8b           \n"  // db - (db * a / 256)
+    "uqsub      v5.8b, v5.8b, v17.8b           \n"  // dg - (dg * a / 256)
+    "uqsub      v6.8b, v6.8b, v18.8b           \n"  // dr - (dr * a / 256)
+    "uqadd      v0.8b, v0.8b, v4.8b            \n"  // + sb
+    "uqadd      v1.8b, v1.8b, v5.8b            \n"  // + sg
+    "uqadd      v2.8b, v2.8b, v6.8b            \n"  // + sr
+    "movi       v3.8b, #255                    \n"  // a = 255
     MEMACCESS(2)
-    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 pixels of ARGB.
-    "bge        8b                             \n"
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB pixels
+    "b.ge       8b                             \n"
 
   "89:                                         \n"
-    "adds       %3, #8-1                       \n"
-    "blt        99f                            \n"
+    "adds       %3, %3, #8-1                   \n"
+    "b.lt       99f                            \n"
 
     // Blend 1 pixels.
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d0[0],d1[0],d2[0],d3[0]}, [%0]! \n"  // load 1 pixel ARGB0.
+    "ld4        {v0.b,v1.b,v2.b,v3.b}[0], [%0], #4 \n"  // load 1 pixel ARGB0.
     MEMACCESS(1)
-    "vld4.8     {d4[0],d5[0],d6[0],d7[0]}, [%1]! \n"  // load 1 pixel ARGB1.
+    "ld4        {v4.b,v5.b,v6.b,v7.b}[0], [%1], #4 \n"  // load 1 pixel ARGB1.
     "subs       %3, %3, #1                     \n"  // 1 processed per loop.
-    "vmull.u8   q10, d4, d3                    \n"  // db * a
-    "vmull.u8   q11, d5, d3                    \n"  // dg * a
-    "vmull.u8   q12, d6, d3                    \n"  // dr * a
-    "vqrshrn.u16 d20, q10, #8                  \n"  // db >>= 8
-    "vqrshrn.u16 d21, q11, #8                  \n"  // dg >>= 8
-    "vqrshrn.u16 d22, q12, #8                  \n"  // dr >>= 8
-    "vqsub.u8   q2, q2, q10                    \n"  // dbg - dbg * a / 256
-    "vqsub.u8   d6, d6, d22                    \n"  // dr - dr * a / 256
-    "vqadd.u8   q0, q0, q2                     \n"  // + sbg
-    "vqadd.u8   d2, d2, d6                     \n"  // + sr
-    "vmov.u8    d3, #255                       \n"  // a = 255
+    "umull      v16.8h, v4.8b, v3.8b           \n"  // db * a
+    "umull      v17.8h, v5.8b, v3.8b           \n"  // dg * a
+    "umull      v18.8h, v6.8b, v3.8b           \n"  // dr * a
+    "uqrshrn    v16.8b, v16.8h, #8             \n"  // db >>= 8
+    "uqrshrn    v17.8b, v17.8h, #8             \n"  // dg >>= 8
+    "uqrshrn    v18.8b, v18.8h, #8             \n"  // dr >>= 8
+    "uqsub      v4.8b, v4.8b, v16.8b           \n"  // db - (db * a / 256)
+    "uqsub      v5.8b, v5.8b, v17.8b           \n"  // dg - (dg * a / 256)
+    "uqsub      v6.8b, v6.8b, v18.8b           \n"  // dr - (dr * a / 256)
+    "uqadd      v0.8b, v0.8b, v4.8b            \n"  // + sb
+    "uqadd      v1.8b, v1.8b, v5.8b            \n"  // + sg
+    "uqadd      v2.8b, v2.8b, v6.8b            \n"  // + sr
+    "movi       v3.8b, #255                    \n"  // a = 255
     MEMACCESS(2)
-    "vst4.8     {d0[0],d1[0],d2[0],d3[0]}, [%2]! \n"  // store 1 pixel.
-    "bge        1b                             \n"
+    "st4        {v0.b,v1.b,v2.b,v3.b}[0], [%2], #4 \n"  // store 1 pixel.
+    "b.ge       1b                             \n"
 
   "99:                                         \n"
 
@@ -2777,7 +2538,8 @@ void ARGBBlendRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
     "+r"(dst_argb),     // %2
     "+r"(width)         // %3
   :
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q10", "q11", "q12"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v16", "v17", "v18"
   );
 }
 #endif  // HAS_ARGBBLENDROW_NEON
@@ -2789,22 +2551,22 @@ void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
     // Attenuate 8 pixels.
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ARGB.
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q10, d0, d3                    \n"  // b * a
-    "vmull.u8   q11, d1, d3                    \n"  // g * a
-    "vmull.u8   q12, d2, d3                    \n"  // r * a
-    "vqrshrn.u16 d0, q10, #8                   \n"  // b >>= 8
-    "vqrshrn.u16 d1, q11, #8                   \n"  // g >>= 8
-    "vqrshrn.u16 d2, q12, #8                   \n"  // r >>= 8
-    MEMACCESS(1)
-    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
-    "bgt        1b                             \n"
+    "umull      v4.8h, v0.8b, v3.8b            \n"  // b * a
+    "umull      v5.8h, v1.8b, v3.8b            \n"  // g * a
+    "umull      v6.8h, v2.8b, v3.8b            \n"  // r * a
+    "uqrshrn    v0.8b, v4.8h, #8               \n"  // b >>= 8
+    "uqrshrn    v1.8b, v5.8h, #8               \n"  // g >>= 8
+    "uqrshrn    v2.8b, v6.8h, #8               \n"  // r >>= 8
+    MEMACCESS(1)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
   : "+r"(src_argb),   // %0
     "+r"(dst_argb),   // %1
     "+r"(width)       // %2
   :
-  : "cc", "memory", "q0", "q1", "q10", "q11", "q12"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6"
   );
 }
 #endif  // HAS_ARGBATTENUATEROW_NEON
@@ -2815,41 +2577,40 @@ void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
 void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size,
                           int interval_offset, int width) {
   asm volatile (
-    "vdup.u16   q8, %2                         \n"
-    "vshr.u16   q8, q8, #1                     \n"  // scale >>= 1
-    "vdup.u16   q9, %3                         \n"  // interval multiply.
-    "vdup.u16   q10, %4                        \n"  // interval add
+    "dup        v4.8h, %w2                     \n"
+    "ushr       v4.8h, v4.8h, #1               \n"  // scale >>= 1
+    "dup        v5.8h, %w3                     \n"  // interval multiply.
+    "dup        v6.8h, %w4                     \n"  // interval add
 
     // 8 pixel loop.
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d0, d2, d4, d6}, [%0]         \n"  // load 8 pixels of ARGB.
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0]  \n"  // load 8 pixels of ARGB.
     "subs       %1, %1, #8                     \n"  // 8 processed per loop.
-    "vmovl.u8   q0, d0                         \n"  // b (0 .. 255)
-    "vmovl.u8   q1, d2                         \n"
-    "vmovl.u8   q2, d4                         \n"
-    "vqdmulh.s16 q0, q0, q8                    \n"  // b * scale
-    "vqdmulh.s16 q1, q1, q8                    \n"  // g
-    "vqdmulh.s16 q2, q2, q8                    \n"  // r
-    "vmul.u16   q0, q0, q9                     \n"  // b * interval_size
-    "vmul.u16   q1, q1, q9                     \n"  // g
-    "vmul.u16   q2, q2, q9                     \n"  // r
-    "vadd.u16   q0, q0, q10                    \n"  // b + interval_offset
-    "vadd.u16   q1, q1, q10                    \n"  // g
-    "vadd.u16   q2, q2, q10                    \n"  // r
-    "vqmovn.u16 d0, q0                         \n"
-    "vqmovn.u16 d2, q1                         \n"
-    "vqmovn.u16 d4, q2                         \n"
-    MEMACCESS(0)
-    "vst4.8     {d0, d2, d4, d6}, [%0]!        \n"  // store 8 pixels of ARGB.
-    "bgt        1b                             \n"
+    "uxtl       v0.8h, v0.8b                   \n"  // b (0 .. 255)
+    "uxtl       v1.8h, v1.8b                   \n"
+    "uxtl       v2.8h, v2.8b                   \n"
+    "sqdmulh    v0.8h, v0.8h, v4.8h            \n"  // b * scale
+    "sqdmulh    v1.8h, v1.8h, v4.8h            \n"  // g
+    "sqdmulh    v2.8h, v2.8h, v4.8h            \n"  // r
+    "mul        v0.8h, v0.8h, v5.8h            \n"  // b * interval_size
+    "mul        v1.8h, v1.8h, v5.8h            \n"  // g
+    "mul        v2.8h, v2.8h, v5.8h            \n"  // r
+    "add        v0.8h, v0.8h, v6.8h            \n"  // b + interval_offset
+    "add        v1.8h, v1.8h, v6.8h            \n"  // g
+    "add        v2.8h, v2.8h, v6.8h            \n"  // r
+    "uqxtn      v0.8b, v0.8h                   \n"
+    "uqxtn      v1.8b, v1.8h                   \n"
+    "uqxtn      v2.8b, v2.8h                   \n"
+    MEMACCESS(0)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
   : "+r"(dst_argb),       // %0
     "+r"(width)           // %1
   : "r"(scale),           // %2
     "r"(interval_size),   // %3
     "r"(interval_offset)  // %4
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "q10"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6"
   );
 }
 #endif  // HAS_ARGBQUANTIZEROW_NEON
@@ -2861,36 +2622,35 @@ void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size,
 void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width,
                        uint32 value) {
   asm volatile (
-    "vdup.u32   q0, %3                         \n"  // duplicate scale value.
-    "vzip.u8    d0, d1                         \n"  // d0 aarrggbb.
-    "vshr.u16   q0, q0, #1                     \n"  // scale / 2.
+    "dup        v0.4s, %w3                     \n"  // duplicate scale value.
+    "zip1       v0.8b, v0.8b, v0.8b            \n"  // v0.8b aarrggbb.
+    "ushr       v0.8h, v0.8h, #1               \n"  // scale / 2.
 
     // 8 pixel loop.
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d20, d22, d24, d26}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vmovl.u8   q10, d20                       \n"  // b (0 .. 255)
-    "vmovl.u8   q11, d22                       \n"
-    "vmovl.u8   q12, d24                       \n"
-    "vmovl.u8   q13, d26                       \n"
-    "vqrdmulh.s16 q10, q10, d0[0]              \n"  // b * scale * 2
-    "vqrdmulh.s16 q11, q11, d0[1]              \n"  // g
-    "vqrdmulh.s16 q12, q12, d0[2]              \n"  // r
-    "vqrdmulh.s16 q13, q13, d0[3]              \n"  // a
-    "vqmovn.u16 d20, q10                       \n"
-    "vqmovn.u16 d22, q11                       \n"
-    "vqmovn.u16 d24, q12                       \n"
-    "vqmovn.u16 d26, q13                       \n"
-    MEMACCESS(1)
-    "vst4.8     {d20, d22, d24, d26}, [%1]!    \n"  // store 8 pixels of ARGB.
-    "bgt        1b                             \n"
+    "uxtl       v4.8h, v4.8b                   \n"  // b (0 .. 255)
+    "uxtl       v5.8h, v5.8b                   \n"
+    "uxtl       v6.8h, v6.8b                   \n"
+    "uxtl       v7.8h, v7.8b                   \n"
+    "sqrdmulh   v4.8h, v4.8h, v0.h[0]          \n"  // b * scale * 2
+    "sqrdmulh   v5.8h, v5.8h, v0.h[1]          \n"  // g
+    "sqrdmulh   v6.8h, v6.8h, v0.h[2]          \n"  // r
+    "sqrdmulh   v7.8h, v7.8h, v0.h[3]          \n"  // a
+    "uqxtn      v4.8b, v4.8h                   \n"
+    "uqxtn      v5.8b, v5.8h                   \n"
+    "uqxtn      v6.8b, v6.8h                   \n"
+    "uqxtn      v7.8b, v7.8h                   \n"
+    MEMACCESS(1)
+    "st4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
   : "+r"(src_argb),       // %0
     "+r"(dst_argb),       // %1
     "+r"(width)           // %2
   : "r"(value)            // %3
-  : "cc", "memory", "q0", "q10", "q11", "q12", "q13"
+  : "cc", "memory", "v0", "v4", "v5", "v6", "v7"
   );
 }
 #endif  // HAS_ARGBSHADEROW_NEON
@@ -2901,28 +2661,27 @@ void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width,
 #ifdef HAS_ARGBGRAYROW_NEON
 void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
   asm volatile (
-    "vmov.u8    d24, #15                       \n"  // B * 0.11400 coefficient
-    "vmov.u8    d25, #75                       \n"  // G * 0.58700 coefficient
-    "vmov.u8    d26, #38                       \n"  // R * 0.29900 coefficient
-    ".p2align   2                              \n"
+    "movi       v24.8b, #15                    \n"  // B * 0.11400 coefficient
+    "movi       v25.8b, #75                    \n"  // G * 0.58700 coefficient
+    "movi       v26.8b, #38                    \n"  // R * 0.29900 coefficient
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q2, d0, d24                    \n"  // B
-    "vmlal.u8   q2, d1, d25                    \n"  // G
-    "vmlal.u8   q2, d2, d26                    \n"  // R
-    "vqrshrun.s16 d0, q2, #7                   \n"  // 15 bit to 8 bit B
-    "vmov       d1, d0                         \n"  // G
-    "vmov       d2, d0                         \n"  // R
-    MEMACCESS(1)
-    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 ARGB pixels.
-    "bgt        1b                             \n"
+    "umull      v4.8h, v0.8b, v24.8b           \n"  // B
+    "umlal      v4.8h, v1.8b, v25.8b           \n"  // G
+    "umlal      v4.8h, v2.8b, v26.8b           \n"  // R
+    "sqrshrun   v0.8b, v4.8h, #7               \n"  // 15 bit to 8 bit B
+    "orr        v1.8b, v0.8b, v0.8b            \n"  // G
+    "orr        v2.8b, v0.8b, v0.8b            \n"  // R
+    MEMACCESS(1)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 pixels.
+    "b.gt       1b                             \n"
   : "+r"(src_argb),  // %0
     "+r"(dst_argb),  // %1
     "+r"(width)      // %2
   :
-  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v24", "v25", "v26"
   );
 }
 #endif  // HAS_ARGBGRAYROW_NEON
@@ -2935,40 +2694,39 @@ void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
 #ifdef HAS_ARGBSEPIAROW_NEON
 void ARGBSepiaRow_NEON(uint8* dst_argb, int width) {
   asm volatile (
-    "vmov.u8    d20, #17                       \n"  // BB coefficient
-    "vmov.u8    d21, #68                       \n"  // BG coefficient
-    "vmov.u8    d22, #35                       \n"  // BR coefficient
-    "vmov.u8    d24, #22                       \n"  // GB coefficient
-    "vmov.u8    d25, #88                       \n"  // GG coefficient
-    "vmov.u8    d26, #45                       \n"  // GR coefficient
-    "vmov.u8    d28, #24                       \n"  // BB coefficient
-    "vmov.u8    d29, #98                       \n"  // BG coefficient
-    "vmov.u8    d30, #50                       \n"  // BR coefficient
-    ".p2align   2                              \n"
+    "movi       v20.8b, #17                    \n"  // BB coefficient
+    "movi       v21.8b, #68                    \n"  // BG coefficient
+    "movi       v22.8b, #35                    \n"  // BR coefficient
+    "movi       v24.8b, #22                    \n"  // GB coefficient
+    "movi       v25.8b, #88                    \n"  // GG coefficient
+    "movi       v26.8b, #45                    \n"  // GR coefficient
+    "movi       v28.8b, #24                    \n"  // BB coefficient
+    "movi       v29.8b, #98                    \n"  // BG coefficient
+    "movi       v30.8b, #50                    \n"  // BR coefficient
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]         \n"  // load 8 ARGB pixels.
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0] \n"  // load 8 ARGB pixels.
     "subs       %1, %1, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q2, d0, d20                    \n"  // B to Sepia B
-    "vmlal.u8   q2, d1, d21                    \n"  // G
-    "vmlal.u8   q2, d2, d22                    \n"  // R
-    "vmull.u8   q3, d0, d24                    \n"  // B to Sepia G
-    "vmlal.u8   q3, d1, d25                    \n"  // G
-    "vmlal.u8   q3, d2, d26                    \n"  // R
-    "vmull.u8   q8, d0, d28                    \n"  // B to Sepia R
-    "vmlal.u8   q8, d1, d29                    \n"  // G
-    "vmlal.u8   q8, d2, d30                    \n"  // R
-    "vqshrn.u16 d0, q2, #7                     \n"  // 16 bit to 8 bit B
-    "vqshrn.u16 d1, q3, #7                     \n"  // 16 bit to 8 bit G
-    "vqshrn.u16 d2, q8, #7                     \n"  // 16 bit to 8 bit R
-    MEMACCESS(0)
-    "vst4.8     {d0, d1, d2, d3}, [%0]!        \n"  // store 8 ARGB pixels.
-    "bgt        1b                             \n"
+    "umull      v4.8h, v0.8b, v20.8b           \n"  // B to Sepia B
+    "umlal      v4.8h, v1.8b, v21.8b           \n"  // G
+    "umlal      v4.8h, v2.8b, v22.8b           \n"  // R
+    "umull      v5.8h, v0.8b, v24.8b           \n"  // B to Sepia G
+    "umlal      v5.8h, v1.8b, v25.8b           \n"  // G
+    "umlal      v5.8h, v2.8b, v26.8b           \n"  // R
+    "umull      v6.8h, v0.8b, v28.8b           \n"  // B to Sepia R
+    "umlal      v6.8h, v1.8b, v29.8b           \n"  // G
+    "umlal      v6.8h, v2.8b, v30.8b           \n"  // R
+    "uqshrn     v0.8b, v4.8h, #7               \n"  // 16 bit to 8 bit B
+    "uqshrn     v1.8b, v5.8h, #7               \n"  // 16 bit to 8 bit G
+    "uqshrn     v2.8b, v6.8h, #7               \n"  // 16 bit to 8 bit R
+    MEMACCESS(0)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // store 8 pixels.
+    "b.gt       1b                             \n"
   : "+r"(dst_argb),  // %0
     "+r"(width)      // %1
   :
-  : "cc", "memory", "q0", "q1", "q2", "q3",
-    "q10", "q11", "q12", "q13", "q14", "q15"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v24", "v25", "v26", "v28", "v29", "v30"
   );
 }
 #endif  // HAS_ARGBSEPIAROW_NEON
@@ -2981,60 +2739,59 @@ void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb,
                              const int8* matrix_argb, int width) {
   asm volatile (
     MEMACCESS(3)
-    "vld1.8     {q2}, [%3]                     \n"  // load 3 ARGB vectors.
-    "vmovl.s8   q0, d4                         \n"  // B,G coefficients s16.
-    "vmovl.s8   q1, d5                         \n"  // R,A coefficients s16.
+    "ld1        {v2.16b}, [%3]                 \n"  // load 3 ARGB vectors.
+    "sxtl       v0.8h, v2.8b                   \n"  // B,G coefficients s16.
+    "sxtl2      v1.8h, v2.16b                  \n"  // R,A coefficients s16.
 
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d16, d18, d20, d22}, [%0]!    \n"  // load 8 ARGB pixels.
+    "ld4        {v16.8b,v17.8b,v18.8b,v19.8b}, [%0], #32 \n"  // load 8 pixels.
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vmovl.u8   q8, d16                        \n"  // b (0 .. 255) 16 bit
-    "vmovl.u8   q9, d18                        \n"  // g
-    "vmovl.u8   q10, d20                       \n"  // r
-    "vmovl.u8   q15, d22                       \n"  // a
-    "vmul.s16   q12, q8, d0[0]                 \n"  // B = B * Matrix B
-    "vmul.s16   q13, q8, d1[0]                 \n"  // G = B * Matrix G
-    "vmul.s16   q14, q8, d2[0]                 \n"  // R = B * Matrix R
-    "vmul.s16   q15, q8, d3[0]                 \n"  // A = B * Matrix A
-    "vmul.s16   q4, q9, d0[1]                  \n"  // B += G * Matrix B
-    "vmul.s16   q5, q9, d1[1]                  \n"  // G += G * Matrix G
-    "vmul.s16   q6, q9, d2[1]                  \n"  // R += G * Matrix R
-    "vmul.s16   q7, q9, d3[1]                  \n"  // A += G * Matrix A
-    "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
-    "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
-    "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
-    "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
-    "vmul.s16   q4, q10, d0[2]                 \n"  // B += R * Matrix B
-    "vmul.s16   q5, q10, d1[2]                 \n"  // G += R * Matrix G
-    "vmul.s16   q6, q10, d2[2]                 \n"  // R += R * Matrix R
-    "vmul.s16   q7, q10, d3[2]                 \n"  // A += R * Matrix A
-    "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
-    "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
-    "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
-    "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
-    "vmul.s16   q4, q15, d0[3]                 \n"  // B += A * Matrix B
-    "vmul.s16   q5, q15, d1[3]                 \n"  // G += A * Matrix G
-    "vmul.s16   q6, q15, d2[3]                 \n"  // R += A * Matrix R
-    "vmul.s16   q7, q15, d3[3]                 \n"  // A += A * Matrix A
-    "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
-    "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
-    "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
-    "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
-    "vqshrun.s16 d16, q12, #6                  \n"  // 16 bit to 8 bit B
-    "vqshrun.s16 d18, q13, #6                  \n"  // 16 bit to 8 bit G
-    "vqshrun.s16 d20, q14, #6                  \n"  // 16 bit to 8 bit R
-    "vqshrun.s16 d22, q15, #6                  \n"  // 16 bit to 8 bit A
-    MEMACCESS(1)
-    "vst4.8     {d16, d18, d20, d22}, [%1]!    \n"  // store 8 ARGB pixels.
-    "bgt        1b                             \n"
+    "uxtl       v16.8h, v16.8b                 \n"  // b (0 .. 255) 16 bit
+    "uxtl       v17.8h, v17.8b                 \n"  // g
+    "uxtl       v18.8h, v18.8b                 \n"  // r
+    "uxtl       v19.8h, v19.8b                 \n"  // a
+    "mul        v22.8h, v16.8h, v0.h[0]        \n"  // B = B * Matrix B
+    "mul        v23.8h, v16.8h, v0.h[4]        \n"  // G = B * Matrix G
+    "mul        v24.8h, v16.8h, v1.h[0]        \n"  // R = B * Matrix R
+    "mul        v25.8h, v16.8h, v1.h[4]        \n"  // A = B * Matrix A
+    "mul        v4.8h, v17.8h, v0.h[1]         \n"  // B += G * Matrix B
+    "mul        v5.8h, v17.8h, v0.h[5]         \n"  // G += G * Matrix G
+    "mul        v6.8h, v17.8h, v1.h[1]         \n"  // R += G * Matrix R
+    "mul        v7.8h, v17.8h, v1.h[5]         \n"  // A += G * Matrix A
+    "sqadd      v22.8h, v22.8h, v4.8h          \n"  // Accumulate B
+    "sqadd      v23.8h, v23.8h, v5.8h          \n"  // Accumulate G
+    "sqadd      v24.8h, v24.8h, v6.8h          \n"  // Accumulate R
+    "sqadd      v25.8h, v25.8h, v7.8h          \n"  // Accumulate A
+    "mul        v4.8h, v18.8h, v0.h[2]         \n"  // B += R * Matrix B
+    "mul        v5.8h, v18.8h, v0.h[6]         \n"  // G += R * Matrix G
+    "mul        v6.8h, v18.8h, v1.h[2]         \n"  // R += R * Matrix R
+    "mul        v7.8h, v18.8h, v1.h[6]         \n"  // A += R * Matrix A
+    "sqadd      v22.8h, v22.8h, v4.8h          \n"  // Accumulate B
+    "sqadd      v23.8h, v23.8h, v5.8h          \n"  // Accumulate G
+    "sqadd      v24.8h, v24.8h, v6.8h          \n"  // Accumulate R
+    "sqadd      v25.8h, v25.8h, v7.8h          \n"  // Accumulate A
+    "mul        v4.8h, v19.8h, v0.h[3]         \n"  // B += A * Matrix B
+    "mul        v5.8h, v19.8h, v0.h[7]         \n"  // G += A * Matrix G
+    "mul        v6.8h, v19.8h, v1.h[3]         \n"  // R += A * Matrix R
+    "mul        v7.8h, v19.8h, v1.h[7]         \n"  // A += A * Matrix A
+    "sqadd      v22.8h, v22.8h, v4.8h          \n"  // Accumulate B
+    "sqadd      v23.8h, v23.8h, v5.8h          \n"  // Accumulate G
+    "sqadd      v24.8h, v24.8h, v6.8h          \n"  // Accumulate R
+    "sqadd      v25.8h, v25.8h, v7.8h          \n"  // Accumulate A
+    "sqshrun    v16.8b, v22.8h, #6             \n"  // 16 bit to 8 bit B
+    "sqshrun    v17.8b, v23.8h, #6             \n"  // 16 bit to 8 bit G
+    "sqshrun    v18.8b, v24.8h, #6             \n"  // 16 bit to 8 bit R
+    "sqshrun    v19.8b, v25.8h, #6             \n"  // 16 bit to 8 bit A
+    MEMACCESS(1)
+    "st4        {v16.8b,v17.8b,v18.8b,v19.8b}, [%1], #32 \n"  // store 8 pixels.
+    "b.gt       1b                             \n"
   : "+r"(src_argb),   // %0
     "+r"(dst_argb),   // %1
     "+r"(width)       // %2
   : "r"(matrix_argb)  // %3
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9",
-    "q10", "q11", "q12", "q13", "q14", "q15"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17",
+    "v18", "v19", "v22", "v23", "v24", "v25"
   );
 }
 #endif  // HAS_ARGBCOLORMATRIXROW_NEON
@@ -3046,12 +2803,11 @@ void ARGBMultiplyRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
                           uint8* dst_argb, int width) {
   asm volatile (
     // 8 pixel loop.
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "ld4        {v0.8b-v3.8b}, [%0], #32       \n"  // load 8 ARGB pixels.
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
     MEMACCESS(1)
-    "ld4        {v4.8b-v7.8b}, [%1], #32       \n"  // load 8 more ARGB pixels.
+    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 more pixels.
     "subs       %3, %3, #8                     \n"  // 8 processed per loop.
     "umull      v0.8h, v0.8b, v4.8b            \n"  // multiply B
     "umull      v1.8h, v1.8b, v5.8b            \n"  // multiply G
@@ -3062,8 +2818,8 @@ void ARGBMultiplyRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
     "rshrn      v2.8b, v2.8h, #8               \n"  // 16 bit to 8 bit R
     "rshrn      v3.8b, v3.8h, #8               \n"  // 16 bit to 8 bit A
     MEMACCESS(2)
-    "st4        {v0.8b-v3.8b}, [%2], #32       \n"  // store 8 ARGB pixels.
-    "bgt        1b                             \n"
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
 
   : "+r"(src_argb0),  // %0
     "+r"(src_argb1),  // %1
@@ -3081,20 +2837,19 @@ void ARGBAddRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
                      uint8* dst_argb, int width) {
   asm volatile (
     // 8 pixel loop.
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "ld4        {v0.8b-v3.8b}, [%0], #32       \n"  // load 8 ARGB pixels.
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
     MEMACCESS(1)
-    "ld4        {v4.8b-v7.8b}, [%1], #32       \n"  // load 8 more ARGB pixels.
+    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 more pixels.
     "subs       %3, %3, #8                     \n"  // 8 processed per loop.
     "uqadd      v0.8b, v0.8b, v4.8b            \n"
     "uqadd      v1.8b, v1.8b, v5.8b            \n"
     "uqadd      v2.8b, v2.8b, v6.8b            \n"
     "uqadd      v3.8b, v3.8b, v7.8b            \n"
     MEMACCESS(2)
-    "st4        {v0.8b-v3.8b}, [%2], #32       \n"  // store 8 ARGB pixels.
-    "bgt        1b                             \n"
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
 
   : "+r"(src_argb0),  // %0
     "+r"(src_argb1),  // %1
@@ -3112,20 +2867,19 @@ void ARGBSubtractRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
                           uint8* dst_argb, int width) {
   asm volatile (
     // 8 pixel loop.
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "ld4        {v0.8b-v3.8b}, [%0], #32       \n"  // load 8 ARGB pixels.
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
     MEMACCESS(1)
-    "ld4        {v4.8b-v7.8b}, [%1], #32       \n"  // load 8 more ARGB pixels.
+    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 more pixels.
     "subs       %3, %3, #8                     \n"  // 8 processed per loop.
     "uqsub      v0.8b, v0.8b, v4.8b            \n"
     "uqsub      v1.8b, v1.8b, v5.8b            \n"
     "uqsub      v2.8b, v2.8b, v6.8b            \n"
     "uqsub      v3.8b, v3.8b, v7.8b            \n"
     MEMACCESS(2)
-    "st4        {v0.8b-v3.8b}, [%2], #32       \n"  // store 8 ARGB pixels.
-    "bgt        1b                             \n"
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
 
   : "+r"(src_argb0),  // %0
     "+r"(src_argb1),  // %1
@@ -3148,7 +2902,6 @@ void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
   asm volatile (
     "movi       v3.8b, #255                    \n"  // alpha
     // 8 pixel loop.
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
     "ld1        {v0.8b}, [%0], #8              \n"  // load 8 sobelx.
@@ -3156,11 +2909,11 @@ void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
     "ld1        {v1.8b}, [%1], #8              \n"  // load 8 sobely.
     "subs       %3, %3, #8                     \n"  // 8 processed per loop.
     "uqadd      v0.8b, v0.8b, v1.8b            \n"  // add
-    "mov        v1.8b, v0.8b                   \n"
-    "mov        v2.8b, v0.8b                   \n"
+    "orr        v1.8b, v0.8b, v0.8b            \n"
+    "orr        v2.8b, v0.8b, v0.8b            \n"
     MEMACCESS(2)
-    "st4        {v0.8b-v3.8b}, [%2], #32       \n"  // store 8 ARGB pixels.
-    "bgt        1b                             \n"
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
   : "+r"(src_sobelx),  // %0
     "+r"(src_sobely),  // %1
     "+r"(dst_argb),    // %2
@@ -3177,7 +2930,6 @@ void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
                           uint8* dst_y, int width) {
   asm volatile (
     // 16 pixel loop.
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
     "ld1        {v0.16b}, [%0], #16            \n"  // load 16 sobelx.
@@ -3187,7 +2939,7 @@ void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
     "uqadd      v0.16b, v0.16b, v1.16b         \n"  // add
     MEMACCESS(2)
     "st1        {v0.16b}, [%2], #16            \n"  // store 16 pixels.
-    "bgt        1b                             \n"
+    "b.gt       1b                             \n"
   : "+r"(src_sobelx),  // %0
     "+r"(src_sobely),  // %1
     "+r"(dst_y),       // %2
@@ -3209,7 +2961,6 @@ void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
   asm volatile (
     "movi       v3.8b, #255                    \n"  // alpha
     // 8 pixel loop.
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
     "ld1        {v2.8b}, [%0], #8              \n"  // load 8 sobelx.
@@ -3218,8 +2969,8 @@ void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
     "subs       %3, %3, #8                     \n"  // 8 processed per loop.
     "uqadd      v1.8b, v0.8b, v2.8b            \n"  // add
     MEMACCESS(2)
-    "st4        {v0.8b-v3.8b}, [%2], #32       \n"  // store 8 ARGB pixels.
-    "bgt        1b                             \n"
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
   : "+r"(src_sobelx),  // %0
     "+r"(src_sobely),  // %1
     "+r"(dst_argb),    // %2
@@ -3238,7 +2989,6 @@ void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
 void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1,
                     const uint8* src_y2, uint8* dst_sobelx, int width) {
   asm volatile (
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
     "ld1        {v0.8b}, [%0],%5               \n"  // top
@@ -3263,7 +3013,7 @@ void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1,
     "uqxtn      v0.8b, v0.8h                   \n"
     MEMACCESS(3)
     "st1        {v0.8b}, [%3], #8              \n"  // store 8 sobelx
-    "bgt        1b                             \n"
+    "b.gt       1b                             \n"
   : "+r"(src_y0),      // %0
     "+r"(src_y1),      // %1
     "+r"(src_y2),      // %2
@@ -3284,7 +3034,6 @@ void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1,
 void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1,
                     uint8* dst_sobely, int width) {
   asm volatile (
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
     "ld1        {v0.8b}, [%0],%4               \n"  // left
@@ -3309,7 +3058,7 @@ void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1,
     "uqxtn      v0.8b, v0.8h                   \n"
     MEMACCESS(2)
     "st1        {v0.8b}, [%2], #8              \n"  // store 8 sobely
-    "bgt        1b                             \n"
+    "b.gt       1b                             \n"
   : "+r"(src_y0),      // %0
     "+r"(src_y1),      // %1
     "+r"(dst_sobely),  // %2
diff --git a/third_party/libyuv/source/row_posix.cc b/third_party/libyuv/source/row_posix.cc
index 106fda568..1a6f7dc4d 100644
--- a/third_party/libyuv/source/row_posix.cc
+++ b/third_party/libyuv/source/row_posix.cc
@@ -1,3 +1,4 @@
+// VERSION 2
 /*
  *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
  *
@@ -92,6 +93,7 @@ static uvec8 kAddY16 = {
   16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u
 };
 
+// 7 bit fixed point 0.5.
 static vec16 kAddYJ64 = {
   64, 64, 64, 64, 64, 64, 64, 64
 };
@@ -221,7 +223,7 @@ void TestRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix) {
   "1:                                          \n"
     "movq      " MEMACCESS(0) ",%%xmm0         \n"
     "lea       " MEMLEA(0x8,0) ",%0            \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x20,1) ",%1           \n"
     "sub       $0x8,%2                         \n"
     "jg        1b                              \n"
@@ -229,10 +231,7 @@ void TestRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix) {
     "+r"(dst_argb),  // %1
     "+r"(pix)        // %2
   :
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm5"
-#endif
+  : "memory", "cc", "xmm0", "xmm1", "xmm5"
   );
 }
 #endif  // TESTING
@@ -252,37 +251,6 @@ void I400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix) {
     "punpckhwd %%xmm1,%%xmm1                   \n"
     "por       %%xmm5,%%xmm0                   \n"
     "por       %%xmm5,%%xmm1                   \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
-    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
-    "lea       " MEMLEA(0x20,1) ",%1           \n"
-    "sub       $0x8,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_y),     // %0
-    "+r"(dst_argb),  // %1
-    "+r"(pix)        // %2
-  :
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm5"
-#endif
-  );
-}
-
-void I400ToARGBRow_Unaligned_SSE2(const uint8* src_y, uint8* dst_argb,
-                                  int pix) {
-  asm volatile (
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "pslld     $0x18,%%xmm5                    \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movq      " MEMACCESS(0) ",%%xmm0         \n"
-    "lea       " MEMLEA(0x8,0) ",%0            \n"
-    "punpcklbw %%xmm0,%%xmm0                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpcklwd %%xmm0,%%xmm0                   \n"
-    "punpckhwd %%xmm1,%%xmm1                   \n"
-    "por       %%xmm5,%%xmm0                   \n"
-    "por       %%xmm5,%%xmm1                   \n"
     "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
     "lea       " MEMLEA(0x20,1) ",%1           \n"
@@ -291,11 +259,7 @@ void I400ToARGBRow_Unaligned_SSE2(const uint8* src_y, uint8* dst_argb,
   : "+r"(src_y),     // %0
     "+r"(dst_argb),  // %1
     "+r"(pix)        // %2
-  :
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm5"
-#endif
+  :: "memory", "cc", "xmm0", "xmm1", "xmm5"
   );
 }
 #endif  // HAS_I400TOARGBROW_SSE2
@@ -318,27 +282,24 @@ void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix) {
     "por       %%xmm5,%%xmm2                   \n"
     "palignr   $0xc,%%xmm0,%%xmm1              \n"
     "pshufb    %%xmm4,%%xmm0                   \n"
-    "movdqa    %%xmm2," MEMACCESS2(0x20,1) "   \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x20,1) "   \n"
     "por       %%xmm5,%%xmm0                   \n"
     "pshufb    %%xmm4,%%xmm1                   \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "por       %%xmm5,%%xmm1                   \n"
     "palignr   $0x4,%%xmm3,%%xmm3              \n"
     "pshufb    %%xmm4,%%xmm3                   \n"
-    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
     "por       %%xmm5,%%xmm3                   \n"
-    "sub       $0x10,%2                        \n"
-    "movdqa    %%xmm3," MEMACCESS2(0x30,1) "   \n"
+    "movdqu    %%xmm3," MEMACCESS2(0x30,1) "   \n"
     "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
     "jg        1b                              \n"
   : "+r"(src_rgb24),  // %0
     "+r"(dst_argb),  // %1
     "+r"(pix)        // %2
   : "m"(kShuffleMaskRGB24ToARGB)  // %3
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
+  : "memory", "cc" , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
 }
 
@@ -359,27 +320,24 @@ void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb, int pix) {
     "por       %%xmm5,%%xmm2                   \n"
     "palignr   $0xc,%%xmm0,%%xmm1              \n"
     "pshufb    %%xmm4,%%xmm0                   \n"
-    "movdqa    %%xmm2," MEMACCESS2(0x20,1) "   \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x20,1) "   \n"
     "por       %%xmm5,%%xmm0                   \n"
     "pshufb    %%xmm4,%%xmm1                   \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "por       %%xmm5,%%xmm1                   \n"
     "palignr   $0x4,%%xmm3,%%xmm3              \n"
     "pshufb    %%xmm4,%%xmm3                   \n"
-    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
     "por       %%xmm5,%%xmm3                   \n"
-    "sub       $0x10,%2                        \n"
-    "movdqa    %%xmm3," MEMACCESS2(0x30,1) "   \n"
+    "movdqu    %%xmm3," MEMACCESS2(0x30,1) "   \n"
     "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
     "jg        1b                              \n"
   : "+r"(src_raw),   // %0
     "+r"(dst_argb),  // %1
     "+r"(pix)        // %2
   : "m"(kShuffleMaskRAWToARGB)  // %3
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
 }
 
@@ -417,9 +375,8 @@ void RGB565ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) {
     "movdqa    %%xmm1,%%xmm2                   \n"
     "punpcklbw %%xmm0,%%xmm1                   \n"
     "punpckhbw %%xmm0,%%xmm2                   \n"
-    BUNDLEALIGN
-    MEMOPMEM(movdqa,xmm1,0x00,1,0,2)           //  movdqa  %%xmm1,(%1,%0,2)
-    MEMOPMEM(movdqa,xmm2,0x10,1,0,2)           //  movdqa  %%xmm2,0x10(%1,%0,2)
+    MEMOPMEM(movdqu,xmm1,0x00,1,0,2)           //  movdqu  %%xmm1,(%1,%0,2)
+    MEMOPMEM(movdqu,xmm2,0x10,1,0,2)           //  movdqu  %%xmm2,0x10(%1,%0,2)
     "lea       " MEMLEA(0x10,0) ",%0           \n"
     "sub       $0x8,%2                         \n"
     "jg        1b                              \n"
@@ -427,13 +384,8 @@ void RGB565ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) {
     "+r"(dst),  // %1
     "+r"(pix)   // %2
   :
-  : "memory", "cc", "eax"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-#endif
+  : "memory", "cc", "eax", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
   );
 }
 
@@ -474,9 +426,8 @@ void ARGB1555ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) {
     "movdqa    %%xmm1,%%xmm2                   \n"
     "punpcklbw %%xmm0,%%xmm1                   \n"
     "punpckhbw %%xmm0,%%xmm2                   \n"
-    BUNDLEALIGN
-    MEMOPMEM(movdqa,xmm1,0x00,1,0,2)           //  movdqa  %%xmm1,(%1,%0,2)
-    MEMOPMEM(movdqa,xmm2,0x10,1,0,2)           //  movdqa  %%xmm2,0x10(%1,%0,2)
+    MEMOPMEM(movdqu,xmm1,0x00,1,0,2)           //  movdqu  %%xmm1,(%1,%0,2)
+    MEMOPMEM(movdqu,xmm2,0x10,1,0,2)           //  movdqu  %%xmm2,0x10(%1,%0,2)
     "lea       " MEMLEA(0x10,0) ",%0           \n"
     "sub       $0x8,%2                         \n"
     "jg        1b                              \n"
@@ -484,13 +435,8 @@ void ARGB1555ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) {
     "+r"(dst),  // %1
     "+r"(pix)   // %2
   :
-  : "memory", "cc", "eax"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-#endif
+  : "memory", "cc", "eax", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
   );
 }
 
@@ -518,9 +464,8 @@ void ARGB4444ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) {
     "movdqa    %%xmm0,%%xmm1                   \n"
     "punpcklbw %%xmm2,%%xmm0                   \n"
     "punpckhbw %%xmm2,%%xmm1                   \n"
-    BUNDLEALIGN
-    MEMOPMEM(movdqa,xmm0,0x00,1,0,2)           //  movdqa  %%xmm0,(%1,%0,2)
-    MEMOPMEM(movdqa,xmm1,0x10,1,0,2)           //  movdqa  %%xmm1,0x10(%1,%0,2)
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,2)           //  movdqu  %%xmm0,(%1,%0,2)
+    MEMOPMEM(movdqu,xmm1,0x10,1,0,2)           //  movdqu  %%xmm1,0x10(%1,%0,2)
     "lea       " MEMLEA(0x10,0) ",%0           \n"
     "sub       $0x8,%2                         \n"
     "jg        1b                              \n"
@@ -528,13 +473,8 @@ void ARGB4444ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) {
     "+r"(dst),  // %1
     "+r"(pix)   // %2
   :
-  : "memory", "cc", "eax"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
+  : "memory", "cc", "eax", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
 }
 
@@ -572,10 +512,7 @@ void ARGBToRGB24Row_SSSE3(const uint8* src, uint8* dst, int pix) {
     "+r"(dst),  // %1
     "+r"(pix)   // %2
   : "m"(kShuffleMaskARGBToRGB24)  // %3
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
-#endif
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
   );
 }
 
@@ -613,10 +550,7 @@ void ARGBToRAWRow_SSSE3(const uint8* src, uint8* dst, int pix) {
     "+r"(dst),  // %1
     "+r"(pix)   // %2
   : "m"(kShuffleMaskARGBToRAW)  // %3
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
-#endif
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
   );
 }
 
@@ -631,7 +565,7 @@ void ARGBToRGB565Row_SSE2(const uint8* src, uint8* dst, int pix) {
     "pslld     $0xb,%%xmm5                     \n"
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
     "movdqa    %%xmm0,%%xmm1                   \n"
     "movdqa    %%xmm0,%%xmm2                   \n"
     "pslld     $0x8,%%xmm0                     \n"
@@ -652,11 +586,7 @@ void ARGBToRGB565Row_SSE2(const uint8* src, uint8* dst, int pix) {
   : "+r"(src),  // %0
     "+r"(dst),  // %1
     "+r"(pix)   // %2
-  :
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
+  :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
 }
 
@@ -672,7 +602,7 @@ void ARGBToARGB1555Row_SSE2(const uint8* src, uint8* dst, int pix) {
     "pslld     $0xf,%%xmm7                     \n"
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
     "movdqa    %%xmm0,%%xmm1                   \n"
     "movdqa    %%xmm0,%%xmm2                   \n"
     "movdqa    %%xmm0,%%xmm3                   \n"
@@ -690,17 +620,14 @@ void ARGBToARGB1555Row_SSE2(const uint8* src, uint8* dst, int pix) {
     "packssdw  %%xmm0,%%xmm0                   \n"
     "lea       " MEMLEA(0x10,0) ",%0           \n"
     "movq      %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMACCESS2(0x8,1) ",%1        \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
     "sub       $0x4,%2                         \n"
     "jg        1b                              \n"
   : "+r"(src),  // %0
     "+r"(dst),  // %1
     "+r"(pix)   // %2
-  :
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-#endif
+  :: "memory", "cc",
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
   );
 }
 
@@ -712,7 +639,7 @@ void ARGBToARGB4444Row_SSE2(const uint8* src, uint8* dst, int pix) {
     "psrlw     $0x8,%%xmm3                     \n"
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
     "movdqa    %%xmm0,%%xmm1                   \n"
     "pand      %%xmm3,%%xmm0                   \n"
     "pand      %%xmm4,%%xmm1                   \n"
@@ -728,57 +655,17 @@ void ARGBToARGB4444Row_SSE2(const uint8* src, uint8* dst, int pix) {
   : "+r"(src),  // %0
     "+r"(dst),  // %1
     "+r"(pix)   // %2
-  :
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
-#endif
+  :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
   );
 }
 #endif  // HAS_RGB24TOARGBROW_SSSE3
 
 #ifdef HAS_ARGBTOYROW_SSSE3
+// Convert 16 ARGB pixels (64 bytes) to 16 Y values.
 void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
   asm volatile (
-    "movdqa    %4,%%xmm5                       \n"
     "movdqa    %3,%%xmm4                       \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm1                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm4,%%xmm3                   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "phaddw    %%xmm1,%%xmm0                   \n"
-    "phaddw    %%xmm3,%%xmm2                   \n"
-    "psrlw     $0x7,%%xmm0                     \n"
-    "psrlw     $0x7,%%xmm2                     \n"
-    "packuswb  %%xmm2,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(pix)        // %2
-  : "m"(kARGBToY),   // %3
-    "m"(kAddY16)     // %4
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
-  );
-}
-
-void ARGBToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
-  asm volatile (
     "movdqa    %4,%%xmm5                       \n"
-    "movdqa    %3,%%xmm4                       \n"
     LABELALIGN
   "1:                                          \n"
     "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
@@ -796,34 +683,33 @@ void ARGBToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
     "psrlw     $0x7,%%xmm2                     \n"
     "packuswb  %%xmm2,%%xmm0                   \n"
     "paddb     %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
     "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
     "jg        1b                              \n"
   : "+r"(src_argb),  // %0
     "+r"(dst_y),     // %1
     "+r"(pix)        // %2
   : "m"(kARGBToY),   // %3
     "m"(kAddY16)     // %4
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
 }
 #endif  // HAS_ARGBTOYROW_SSSE3
 
 #ifdef HAS_ARGBTOYJROW_SSSE3
+// Convert 16 ARGB pixels (64 bytes) to 16 YJ values.
+// Same as ARGBToYRow but different coefficients, no add 16, but do rounding.
 void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
   asm volatile (
     "movdqa    %3,%%xmm4                       \n"
     "movdqa    %4,%%xmm5                       \n"
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
     "pmaddubsw %%xmm4,%%xmm0                   \n"
     "pmaddubsw %%xmm4,%%xmm1                   \n"
     "pmaddubsw %%xmm4,%%xmm2                   \n"
@@ -836,158 +722,131 @@ void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
     "psrlw     $0x7,%%xmm0                     \n"
     "psrlw     $0x7,%%xmm2                     \n"
     "packuswb  %%xmm2,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
     "jg        1b                              \n"
   : "+r"(src_argb),  // %0
     "+r"(dst_y),     // %1
     "+r"(pix)        // %2
   : "m"(kARGBToYJ),  // %3
     "m"(kAddYJ64)    // %4
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
 }
+#endif  // HAS_ARGBTOYJROW_SSSE3
+
+#ifdef HAS_ARGBTOYROW_AVX2
+// vpermd for vphaddw + vpackuswb vpermd.
+static const lvec32 kPermdARGBToY_AVX = {
+  0, 4, 1, 5, 2, 6, 3, 7
+};
 
-void ARGBToYJRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+// Convert 32 ARGB pixels (128 bytes) to 32 Y values.
+void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) {
   asm volatile (
-    "movdqa    %3,%%xmm4                       \n"
-    "movdqa    %4,%%xmm5                       \n"
+    "vbroadcastf128 %3,%%ymm4                  \n"
+    "vbroadcastf128 %4,%%ymm5                  \n"
+    "vmovdqu    %5,%%ymm6                      \n"
     LABELALIGN
   "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm1                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm4,%%xmm3                   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "phaddw    %%xmm1,%%xmm0                   \n"
-    "phaddw    %%xmm3,%%xmm2                   \n"
-    "paddw     %%xmm5,%%xmm0                   \n"
-    "paddw     %%xmm5,%%xmm2                   \n"
-    "psrlw     $0x7,%%xmm0                     \n"
-    "psrlw     $0x7,%%xmm2                     \n"
-    "packuswb  %%xmm2,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
+    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1  \n"
+    "vmovdqu    " MEMACCESS2(0x40,0) ",%%ymm2  \n"
+    "vmovdqu    " MEMACCESS2(0x60,0) ",%%ymm3  \n"
+    "vpmaddubsw %%ymm4,%%ymm0,%%ymm0           \n"
+    "vpmaddubsw %%ymm4,%%ymm1,%%ymm1           \n"
+    "vpmaddubsw %%ymm4,%%ymm2,%%ymm2           \n"
+    "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
+    "lea       " MEMLEA(0x80,0) ",%0           \n"
+    "vphaddw    %%ymm1,%%ymm0,%%ymm0           \n"  // mutates.
+    "vphaddw    %%ymm3,%%ymm2,%%ymm2           \n"
+    "vpsrlw     $0x7,%%ymm0,%%ymm0             \n"
+    "vpsrlw     $0x7,%%ymm2,%%ymm2             \n"
+    "vpackuswb  %%ymm2,%%ymm0,%%ymm0           \n"  // mutates.
+    "vpermd     %%ymm0,%%ymm6,%%ymm0           \n"  // unmutate.
+    "vpaddb     %%ymm5,%%ymm0,%%ymm0           \n"  // add 16 for Y
+    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
     "jg        1b                              \n"
+    "vzeroupper                                \n"
   : "+r"(src_argb),  // %0
     "+r"(dst_y),     // %1
     "+r"(pix)        // %2
-  : "m"(kARGBToYJ),  // %3
-    "m"(kAddYJ64)    // %4
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
+  : "m"(kARGBToY),   // %3
+    "m"(kAddY16),    // %4
+    "m"(kPermdARGBToY_AVX)  // %5
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
   );
 }
-#endif  // HAS_ARGBTOYJROW_SSSE3
+#endif  // HAS_ARGBTOYROW_AVX2
 
-#ifdef HAS_ARGBTOUVROW_SSSE3
-// TODO(fbarchard): pass xmm constants to single block of assembly.
-// fpic on GCC 4.2 for OSX runs out of GPR registers. "m" effectively takes
-// 3 registers - ebx, ebp and eax. "m" can be passed with 3 normal registers,
-// or 4 if stack frame is disabled. Doing 2 assembly blocks is a work around
-// and considered unsafe.
-void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                       uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "movdqa    %0,%%xmm4                       \n"
-    "movdqa    %1,%%xmm3                       \n"
-    "movdqa    %2,%%xmm5                       \n"
-  :
-  : "m"(kARGBToU),  // %0
-    "m"(kARGBToV),  // %1
-    "m"(kAddUV128)  // %2
-  );
+#ifdef HAS_ARGBTOYJROW_AVX2
+// Convert 32 ARGB pixels (128 bytes) to 32 Y values.
+void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) {
   asm volatile (
-    "sub       %1,%2                           \n"
+    "vbroadcastf128 %3,%%ymm4                  \n"
+    "vbroadcastf128 %4,%%ymm5                  \n"
+    "vmovdqu    %5,%%ymm6                      \n"
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
-    BUNDLEALIGN
-    MEMOPREG(pavgb,0x00,0,4,1,xmm0)            //  pavgb   (%0,%4,1),%%xmm0
-    MEMOPREG(pavgb,0x10,0,4,1,xmm1)            //  pavgb   0x10(%0,%4,1),%%xmm1
-    MEMOPREG(pavgb,0x20,0,4,1,xmm2)            //  pavgb   0x20(%0,%4,1),%%xmm2
-    MEMOPREG(pavgb,0x30,0,4,1,xmm6)            //  pavgb   0x30(%0,%4,1),%%xmm6
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "movdqa    %%xmm0,%%xmm7                   \n"
-    "shufps    $0x88,%%xmm1,%%xmm0             \n"
-    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
-    "pavgb     %%xmm7,%%xmm0                   \n"
-    "movdqa    %%xmm2,%%xmm7                   \n"
-    "shufps    $0x88,%%xmm6,%%xmm2             \n"
-    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
-    "pavgb     %%xmm7,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "movdqa    %%xmm2,%%xmm6                   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm3,%%xmm1                   \n"
-    "pmaddubsw %%xmm3,%%xmm6                   \n"
-    "phaddw    %%xmm2,%%xmm0                   \n"
-    "phaddw    %%xmm6,%%xmm1                   \n"
-    "psraw     $0x8,%%xmm0                     \n"
-    "psraw     $0x8,%%xmm1                     \n"
-    "packsswb  %%xmm1,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%3                        \n"
-    "movlps    %%xmm0," MEMACCESS(1) "         \n"
-    BUNDLEALIGN
-    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps    %%xmm0,(%1,%2,1)
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
+    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1  \n"
+    "vmovdqu    " MEMACCESS2(0x40,0) ",%%ymm2  \n"
+    "vmovdqu    " MEMACCESS2(0x60,0) ",%%ymm3  \n"
+    "vpmaddubsw %%ymm4,%%ymm0,%%ymm0           \n"
+    "vpmaddubsw %%ymm4,%%ymm1,%%ymm1           \n"
+    "vpmaddubsw %%ymm4,%%ymm2,%%ymm2           \n"
+    "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
+    "lea       " MEMLEA(0x80,0) ",%0           \n"
+    "vphaddw    %%ymm1,%%ymm0,%%ymm0           \n"  // mutates.
+    "vphaddw    %%ymm3,%%ymm2,%%ymm2           \n"
+    "vpaddw     %%ymm5,%%ymm0,%%ymm0           \n"  // Add .5 for rounding.
+    "vpaddw     %%ymm5,%%ymm2,%%ymm2           \n"
+    "vpsrlw     $0x7,%%ymm0,%%ymm0             \n"
+    "vpsrlw     $0x7,%%ymm2,%%ymm2             \n"
+    "vpackuswb  %%ymm2,%%ymm0,%%ymm0           \n"  // mutates.
+    "vpermd     %%ymm0,%%ymm6,%%ymm0           \n"  // unmutate.
+    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
     "jg        1b                              \n"
-  : "+r"(src_argb0),       // %0
-    "+r"(dst_u),           // %1
-    "+r"(dst_v),           // %2
-    "+rm"(width)           // %3
-  : "r"((intptr_t)(src_stride_argb)) // %4
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
-#endif
+    "vzeroupper                                \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kARGBToYJ),   // %3
+    "m"(kAddYJ64),    // %4
+    "m"(kPermdARGBToY_AVX)  // %5
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
   );
 }
+#endif  // HAS_ARGBTOYJROW_AVX2
 
-// TODO(fbarchard): Share code with ARGBToUVRow_SSSE3.
-void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                        uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "movdqa    %0,%%xmm4                       \n"
-    "movdqa    %1,%%xmm3                       \n"
-    "movdqa    %2,%%xmm5                       \n"
-  :
-  : "m"(kARGBToUJ),  // %0
-    "m"(kARGBToVJ),  // %1
-    "m"(kAddUVJ128)  // %2
-  );
+#ifdef HAS_ARGBTOUVROW_SSSE3
+void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
   asm volatile (
+    "movdqa    %5,%%xmm3                       \n"
+    "movdqa    %6,%%xmm4                       \n"
+    "movdqa    %7,%%xmm5                       \n"
     "sub       %1,%2                           \n"
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
-    BUNDLEALIGN
-    MEMOPREG(pavgb,0x00,0,4,1,xmm0)            //  pavgb   (%0,%4,1),%%xmm0
-    MEMOPREG(pavgb,0x10,0,4,1,xmm1)            //  pavgb   0x10(%0,%4,1),%%xmm1
-    MEMOPREG(pavgb,0x20,0,4,1,xmm2)            //  pavgb   0x20(%0,%4,1),%%xmm2
-    MEMOPREG(pavgb,0x30,0,4,1,xmm6)            //  pavgb   0x30(%0,%4,1),%%xmm6
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)            //  movdqu (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)            //  movdqu 0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)            //  movdqu 0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)            //  movdqu 0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+
     "lea       " MEMLEA(0x40,0) ",%0           \n"
     "movdqa    %%xmm0,%%xmm7                   \n"
     "shufps    $0x88,%%xmm1,%%xmm0             \n"
@@ -1005,130 +864,118 @@ void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
     "pmaddubsw %%xmm3,%%xmm6                   \n"
     "phaddw    %%xmm2,%%xmm0                   \n"
     "phaddw    %%xmm6,%%xmm1                   \n"
-    "paddw     %%xmm5,%%xmm0                   \n"
-    "paddw     %%xmm5,%%xmm1                   \n"
     "psraw     $0x8,%%xmm0                     \n"
     "psraw     $0x8,%%xmm1                     \n"
     "packsswb  %%xmm1,%%xmm0                   \n"
-    "sub       $0x10,%3                        \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
     "movlps    %%xmm0," MEMACCESS(1) "         \n"
-    BUNDLEALIGN
-    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps    %%xmm0,(%1,%2,1)
     "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
     "jg        1b                              \n"
   : "+r"(src_argb0),       // %0
     "+r"(dst_u),           // %1
     "+r"(dst_v),           // %2
     "+rm"(width)           // %3
-  : "r"((intptr_t)(src_stride_argb)) // %4
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
-#endif
+  : "r"((intptr_t)(src_stride_argb)), // %4
+    "m"(kARGBToV),  // %5
+    "m"(kARGBToU),  // %6
+    "m"(kAddUV128)  // %7
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
   );
 }
+#endif  // HAS_ARGBTOUVROW_SSSE3
 
-void ARGBToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                                 uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "movdqa    %0,%%xmm4                       \n"
-    "movdqa    %1,%%xmm3                       \n"
-    "movdqa    %2,%%xmm5                       \n"
-  :
-  : "m"(kARGBToU),         // %0
-    "m"(kARGBToV),         // %1
-    "m"(kAddUV128)         // %2
-  );
+#ifdef HAS_ARGBTOUVROW_AVX2
+// vpshufb for vphaddw + vpackuswb packed to shorts.
+static const lvec8 kShufARGBToUV_AVX = {
+  0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15,
+  0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15
+};
+void ARGBToUVRow_AVX2(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) {
   asm volatile (
+    "vbroadcastf128 %5,%%ymm5                  \n"
+    "vbroadcastf128 %6,%%ymm6                  \n"
+    "vbroadcastf128 %7,%%ymm7                  \n"
     "sub       %1,%2                           \n"
     LABELALIGN
   "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
-    BUNDLEALIGN
-    MEMOPREG(movdqu,0x00,0,4,1,xmm7)           //  movdqu  (%0,%4,1),%%xmm7
-    "pavgb     %%xmm7,%%xmm0                   \n"
-    MEMOPREG(movdqu,0x10,0,4,1,xmm7)           //  movdqu  0x10(%0,%4,1),%%xmm7
-    "pavgb     %%xmm7,%%xmm1                   \n"
-    MEMOPREG(movdqu,0x20,0,4,1,xmm7)           //  movdqu  0x20(%0,%4,1),%%xmm7
-    "pavgb     %%xmm7,%%xmm2                   \n"
-    MEMOPREG(movdqu,0x30,0,4,1,xmm7)           //  movdqu  0x30(%0,%4,1),%%xmm7
-    "pavgb     %%xmm7,%%xmm6                   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "movdqa    %%xmm0,%%xmm7                   \n"
-    "shufps    $0x88,%%xmm1,%%xmm0             \n"
-    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
-    "pavgb     %%xmm7,%%xmm0                   \n"
-    "movdqa    %%xmm2,%%xmm7                   \n"
-    "shufps    $0x88,%%xmm6,%%xmm2             \n"
-    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
-    "pavgb     %%xmm7,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "movdqa    %%xmm2,%%xmm6                   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm3,%%xmm1                   \n"
-    "pmaddubsw %%xmm3,%%xmm6                   \n"
-    "phaddw    %%xmm2,%%xmm0                   \n"
-    "phaddw    %%xmm6,%%xmm1                   \n"
-    "psraw     $0x8,%%xmm0                     \n"
-    "psraw     $0x8,%%xmm1                     \n"
-    "packsswb  %%xmm1,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%3                        \n"
-    "movlps    %%xmm0," MEMACCESS(1) "         \n"
-    BUNDLEALIGN
-    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
+    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1  \n"
+    "vmovdqu    " MEMACCESS2(0x40,0) ",%%ymm2  \n"
+    "vmovdqu    " MEMACCESS2(0x60,0) ",%%ymm3  \n"
+    VMEMOPREG(vpavgb,0x00,0,4,1,ymm0,ymm0)     // vpavgb (%0,%4,1),%%ymm0,%%ymm0
+    VMEMOPREG(vpavgb,0x20,0,4,1,ymm1,ymm1)
+    VMEMOPREG(vpavgb,0x40,0,4,1,ymm2,ymm2)
+    VMEMOPREG(vpavgb,0x60,0,4,1,ymm3,ymm3)
+    "lea       " MEMLEA(0x80,0) ",%0           \n"
+    "vshufps    $0x88,%%ymm1,%%ymm0,%%ymm4     \n"
+    "vshufps    $0xdd,%%ymm1,%%ymm0,%%ymm0     \n"
+    "vpavgb     %%ymm4,%%ymm0,%%ymm0           \n"
+    "vshufps    $0x88,%%ymm3,%%ymm2,%%ymm4     \n"
+    "vshufps    $0xdd,%%ymm3,%%ymm2,%%ymm2     \n"
+    "vpavgb     %%ymm4,%%ymm2,%%ymm2           \n"
+
+    "vpmaddubsw %%ymm7,%%ymm0,%%ymm1           \n"
+    "vpmaddubsw %%ymm7,%%ymm2,%%ymm3           \n"
+    "vpmaddubsw %%ymm6,%%ymm0,%%ymm0           \n"
+    "vpmaddubsw %%ymm6,%%ymm2,%%ymm2           \n"
+    "vphaddw    %%ymm3,%%ymm1,%%ymm1           \n"
+    "vphaddw    %%ymm2,%%ymm0,%%ymm0           \n"
+    "vpsraw     $0x8,%%ymm1,%%ymm1             \n"
+    "vpsraw     $0x8,%%ymm0,%%ymm0             \n"
+    "vpacksswb  %%ymm0,%%ymm1,%%ymm0           \n"
+    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
+    "vpshufb    %8,%%ymm0,%%ymm0               \n"
+    "vpaddb     %%ymm5,%%ymm0,%%ymm0           \n"
+
+    "vextractf128 $0x0,%%ymm0," MEMACCESS(1) " \n"
+    VEXTOPMEM(vextractf128,1,ymm0,0x0,1,2,1) // vextractf128 $1,%%ymm0,(%1,%2,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x20,%3                        \n"
     "jg        1b                              \n"
+    "vzeroupper                                \n"
   : "+r"(src_argb0),       // %0
     "+r"(dst_u),           // %1
     "+r"(dst_v),           // %2
     "+rm"(width)           // %3
-  : "r"((intptr_t)(src_stride_argb)) // %4
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
-#endif
+  : "r"((intptr_t)(src_stride_argb)), // %4
+    "m"(kAddUV128),  // %5
+    "m"(kARGBToV),   // %6
+    "m"(kARGBToU),   // %7
+    "m"(kShufARGBToUV_AVX)  // %8
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
   );
 }
+#endif  // HAS_ARGBTOUVROW_AVX2
 
-void ARGBToUVJRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                                  uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "movdqa    %0,%%xmm4                       \n"
-    "movdqa    %1,%%xmm3                       \n"
-    "movdqa    %2,%%xmm5                       \n"
-  :
-  : "m"(kARGBToUJ),         // %0
-    "m"(kARGBToVJ),         // %1
-    "m"(kAddUVJ128)         // %2
-  );
+#ifdef HAS_ARGBTOUVJROW_SSSE3
+// TODO(fbarchard): Share code with ARGBToUVRow_SSSE3.
+void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                        uint8* dst_u, uint8* dst_v, int width) {
   asm volatile (
+    "movdqa    %5,%%xmm3                       \n"
+    "movdqa    %6,%%xmm4                       \n"
+    "movdqa    %7,%%xmm5                       \n"
     "sub       %1,%2                           \n"
     LABELALIGN
   "1:                                          \n"
     "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
-    BUNDLEALIGN
-    MEMOPREG(movdqu,0x00,0,4,1,xmm7)           //  movdqu  (%0,%4,1),%%xmm7
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)            //  movdqu (%0,%4,1),%%xmm7
     "pavgb     %%xmm7,%%xmm0                   \n"
-    MEMOPREG(movdqu,0x10,0,4,1,xmm7)           //  movdqu  0x10(%0,%4,1),%%xmm7
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)            //  movdqu 0x10(%0,%4,1),%%xmm7
     "pavgb     %%xmm7,%%xmm1                   \n"
-    MEMOPREG(movdqu,0x20,0,4,1,xmm7)           //  movdqu  0x20(%0,%4,1),%%xmm7
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)            //  movdqu 0x20(%0,%4,1),%%xmm7
     "pavgb     %%xmm7,%%xmm2                   \n"
-    MEMOPREG(movdqu,0x30,0,4,1,xmm7)           //  movdqu  0x30(%0,%4,1),%%xmm7
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)            //  movdqu 0x30(%0,%4,1),%%xmm7
     "pavgb     %%xmm7,%%xmm6                   \n"
+
     "lea       " MEMLEA(0x40,0) ",%0           \n"
     "movdqa    %%xmm0,%%xmm7                   \n"
     "shufps    $0x88,%%xmm1,%%xmm0             \n"
@@ -1151,104 +998,32 @@ void ARGBToUVJRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
     "psraw     $0x8,%%xmm0                     \n"
     "psraw     $0x8,%%xmm1                     \n"
     "packsswb  %%xmm1,%%xmm0                   \n"
-    "sub       $0x10,%3                        \n"
     "movlps    %%xmm0," MEMACCESS(1) "         \n"
-    BUNDLEALIGN
     MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
     "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
     "jg        1b                              \n"
   : "+r"(src_argb0),       // %0
     "+r"(dst_u),           // %1
     "+r"(dst_v),           // %2
     "+rm"(width)           // %3
-  : "r"((intptr_t)(src_stride_argb))
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
-#endif
+  : "r"((intptr_t)(src_stride_argb)), // %4
+    "m"(kARGBToVJ),  // %5
+    "m"(kARGBToUJ),  // %6
+    "m"(kAddUVJ128)  // %7
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
   );
 }
+#endif  // HAS_ARGBTOUVJROW_SSSE3
 
+#ifdef HAS_ARGBTOUV444ROW_SSSE3
 void ARGBToUV444Row_SSSE3(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
                           int width) {
   asm volatile (
-    "movdqa    %0,%%xmm4                       \n"
-    "movdqa    %1,%%xmm3                       \n"
-    "movdqa    %2,%%xmm5                       \n"
-  :
-  : "m"(kARGBToU),  // %0
-    "m"(kARGBToV),  // %1
-    "m"(kAddUV128)  // %2
-  );
-  asm volatile (
-    "sub       %1,%2                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm1                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm4,%%xmm6                   \n"
-    "phaddw    %%xmm1,%%xmm0                   \n"
-    "phaddw    %%xmm6,%%xmm2                   \n"
-    "psraw     $0x8,%%xmm0                     \n"
-    "psraw     $0x8,%%xmm2                     \n"
-    "packsswb  %%xmm2,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%3                        \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
-    "pmaddubsw %%xmm3,%%xmm0                   \n"
-    "pmaddubsw %%xmm3,%%xmm1                   \n"
-    "pmaddubsw %%xmm3,%%xmm2                   \n"
-    "pmaddubsw %%xmm3,%%xmm6                   \n"
-    "phaddw    %%xmm1,%%xmm0                   \n"
-    "phaddw    %%xmm6,%%xmm2                   \n"
-    "psraw     $0x8,%%xmm0                     \n"
-    "psraw     $0x8,%%xmm2                     \n"
-    "packsswb  %%xmm2,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    BUNDLEALIGN
-    MEMOPMEM(movdqa,xmm0,0x00,1,2,1)           //  movdqa  %%xmm0,(%1,%2,1)
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb),        // %0
-    "+r"(dst_u),           // %1
-    "+r"(dst_v),           // %2
-    "+rm"(width)           // %3
-  :
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm6"
-#endif
-  );
-}
-
-void ARGBToUV444Row_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_u,
-                                    uint8* dst_v, int width) {
-  asm volatile (
-    "movdqa    %0,%%xmm4                       \n"
-    "movdqa    %1,%%xmm3                       \n"
-    "movdqa    %2,%%xmm5                       \n"
-  :
-  : "m"(kARGBToU),  // %0
-    "m"(kARGBToV),  // %1
-    "m"(kAddUV128)  // %2
-  );
-  asm volatile (
+    "movdqa    %4,%%xmm3                       \n"
+    "movdqa    %5,%%xmm4                       \n"
+    "movdqa    %6,%%xmm5                       \n"
     "sub       %1,%2                           \n"
     LABELALIGN
   "1:                                          \n"
@@ -1266,7 +1041,6 @@ void ARGBToUV444Row_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_u,
     "psraw     $0x8,%%xmm2                     \n"
     "packsswb  %%xmm2,%%xmm0                   \n"
     "paddb     %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%3                        \n"
     "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
     "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
@@ -1283,98 +1057,30 @@ void ARGBToUV444Row_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_u,
     "packsswb  %%xmm2,%%xmm0                   \n"
     "paddb     %%xmm5,%%xmm0                   \n"
     "lea       " MEMLEA(0x40,0) ",%0           \n"
-    BUNDLEALIGN
     MEMOPMEM(movdqu,xmm0,0x00,1,2,1)           //  movdqu  %%xmm0,(%1,%2,1)
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%3                        \n"
     "jg        1b                              \n"
   : "+r"(src_argb),        // %0
     "+r"(dst_u),           // %1
     "+r"(dst_v),           // %2
     "+rm"(width)           // %3
-  :
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm6"
-#endif
+  : "m"(kARGBToV),  // %4
+    "m"(kARGBToU),  // %5
+    "m"(kAddUV128)  // %6
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm6"
   );
 }
+#endif  // HAS_ARGBTOUV444ROW_SSSE3
 
+#ifdef HAS_ARGBTOUV422ROW_SSSE3
 void ARGBToUV422Row_SSSE3(const uint8* src_argb0,
                           uint8* dst_u, uint8* dst_v, int width) {
   asm volatile (
-    "movdqa    %0,%%xmm4                       \n"
-    "movdqa    %1,%%xmm3                       \n"
-    "movdqa    %2,%%xmm5                       \n"
-  :
-  : "m"(kARGBToU),  // %0
-    "m"(kARGBToV),  // %1
-    "m"(kAddUV128)  // %2
-  );
-  asm volatile (
-    "sub       %1,%2                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "movdqa    %%xmm0,%%xmm7                   \n"
-    "shufps    $0x88,%%xmm1,%%xmm0             \n"
-    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
-    "pavgb     %%xmm7,%%xmm0                   \n"
-    "movdqa    %%xmm2,%%xmm7                   \n"
-    "shufps    $0x88,%%xmm6,%%xmm2             \n"
-    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
-    "pavgb     %%xmm7,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "movdqa    %%xmm2,%%xmm6                   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm3,%%xmm1                   \n"
-    "pmaddubsw %%xmm3,%%xmm6                   \n"
-    "phaddw    %%xmm2,%%xmm0                   \n"
-    "phaddw    %%xmm6,%%xmm1                   \n"
-    "psraw     $0x8,%%xmm0                     \n"
-    "psraw     $0x8,%%xmm1                     \n"
-    "packsswb  %%xmm1,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%3                        \n"
-    "movlps    %%xmm0," MEMACCESS(1) "         \n"
-    BUNDLEALIGN
-    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb0),       // %0
-    "+r"(dst_u),           // %1
-    "+r"(dst_v),           // %2
-    "+rm"(width)           // %3
-  :
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
-#endif
-  );
-}
-
-void ARGBToUV422Row_Unaligned_SSSE3(const uint8* src_argb0,
-                                    uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "movdqa    %0,%%xmm4                       \n"
-    "movdqa    %1,%%xmm3                       \n"
-    "movdqa    %2,%%xmm5                       \n"
-  :
-  : "m"(kARGBToU),  // %0
-    "m"(kARGBToV),  // %1
-    "m"(kAddUV128)  // %2
-  );
-  asm volatile (
+    "movdqa    %4,%%xmm3                       \n"
+    "movdqa    %5,%%xmm4                       \n"
+    "movdqa    %6,%%xmm5                       \n"
     "sub       %1,%2                           \n"
     LABELALIGN
   "1:                                          \n"
@@ -1403,26 +1109,23 @@ void ARGBToUV422Row_Unaligned_SSSE3(const uint8* src_argb0,
     "psraw     $0x8,%%xmm1                     \n"
     "packsswb  %%xmm1,%%xmm0                   \n"
     "paddb     %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%3                        \n"
     "movlps    %%xmm0," MEMACCESS(1) "         \n"
-    BUNDLEALIGN
     MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
     "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
     "jg        1b                              \n"
   : "+r"(src_argb0),       // %0
     "+r"(dst_u),           // %1
     "+r"(dst_v),           // %2
     "+rm"(width)           // %3
-  :
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
-#endif
+  : "m"(kARGBToV),  // %4
+    "m"(kARGBToU),  // %5
+    "m"(kAddUV128)  // %6
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
   );
 }
+#endif  // HAS_ARGBTOUV422ROW_SSSE3
 
 void BGRAToYRow_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix) {
   asm volatile (
@@ -1430,43 +1133,6 @@ void BGRAToYRow_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix) {
     "movdqa    %3,%%xmm4                       \n"
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm1                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm4,%%xmm3                   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "phaddw    %%xmm1,%%xmm0                   \n"
-    "phaddw    %%xmm3,%%xmm2                   \n"
-    "psrlw     $0x7,%%xmm0                     \n"
-    "psrlw     $0x7,%%xmm2                     \n"
-    "packuswb  %%xmm2,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "jg        1b                              \n"
-  : "+r"(src_bgra),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(pix)        // %2
-  : "m"(kBGRAToY),   // %3
-    "m"(kAddY16)     // %4
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
-  );
-}
-
-void BGRAToYRow_Unaligned_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix) {
-  asm volatile (
-    "movdqa    %4,%%xmm5                       \n"
-    "movdqa    %3,%%xmm4                       \n"
-    LABELALIGN
-  "1:                                          \n"
     "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
     "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
     "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
@@ -1482,116 +1148,41 @@ void BGRAToYRow_Unaligned_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix) {
     "psrlw     $0x7,%%xmm2                     \n"
     "packuswb  %%xmm2,%%xmm0                   \n"
     "paddb     %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
     "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
     "jg        1b                              \n"
   : "+r"(src_bgra),  // %0
     "+r"(dst_y),     // %1
     "+r"(pix)        // %2
   : "m"(kBGRAToY),   // %3
     "m"(kAddY16)     // %4
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
 }
 
 void BGRAToUVRow_SSSE3(const uint8* src_bgra0, int src_stride_bgra,
                        uint8* dst_u, uint8* dst_v, int width) {
   asm volatile (
-    "movdqa    %0,%%xmm4                       \n"
-    "movdqa    %1,%%xmm3                       \n"
-    "movdqa    %2,%%xmm5                       \n"
-  :
-  : "m"(kBGRAToU),         // %0
-    "m"(kBGRAToV),         // %1
-    "m"(kAddUV128)         // %2
-  );
-  asm volatile (
-    "sub       %1,%2                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
-    BUNDLEALIGN
-    MEMOPREG(pavgb,0x00,0,4,1,xmm0)            //  pavgb   (%0,%4,1),%%xmm0
-    MEMOPREG(pavgb,0x10,0,4,1,xmm1)            //  pavgb   0x10(%0,%4,1),%%xmm1
-    MEMOPREG(pavgb,0x20,0,4,1,xmm2)            //  pavgb   0x20(%0,%4,1),%%xmm2
-    MEMOPREG(pavgb,0x30,0,4,1,xmm6)            //  pavgb   0x30(%0,%4,1),%%xmm6
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "movdqa    %%xmm0,%%xmm7                   \n"
-    "shufps    $0x88,%%xmm1,%%xmm0             \n"
-    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
-    "pavgb     %%xmm7,%%xmm0                   \n"
-    "movdqa    %%xmm2,%%xmm7                   \n"
-    "shufps    $0x88,%%xmm6,%%xmm2             \n"
-    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
-    "pavgb     %%xmm7,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "movdqa    %%xmm2,%%xmm6                   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm3,%%xmm1                   \n"
-    "pmaddubsw %%xmm3,%%xmm6                   \n"
-    "phaddw    %%xmm2,%%xmm0                   \n"
-    "phaddw    %%xmm6,%%xmm1                   \n"
-    "psraw     $0x8,%%xmm0                     \n"
-    "psraw     $0x8,%%xmm1                     \n"
-    "packsswb  %%xmm1,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%3                        \n"
-    "movlps    %%xmm0," MEMACCESS(1) "         \n"
-    BUNDLEALIGN
-    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "jg        1b                              \n"
-  : "+r"(src_bgra0),       // %0
-    "+r"(dst_u),           // %1
-    "+r"(dst_v),           // %2
-    "+rm"(width)           // %3
-  : "r"((intptr_t)(src_stride_bgra)) // %4
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
-#endif
-  );
-}
-
-void BGRAToUVRow_Unaligned_SSSE3(const uint8* src_bgra0, int src_stride_bgra,
-                                 uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "movdqa    %0,%%xmm4                       \n"
-    "movdqa    %1,%%xmm3                       \n"
-    "movdqa    %2,%%xmm5                       \n"
-  :
-  : "m"(kBGRAToU),         // %0
-    "m"(kBGRAToV),         // %1
-    "m"(kAddUV128)         // %2
-  );
-  asm volatile (
+    "movdqa    %5,%%xmm3                       \n"
+    "movdqa    %6,%%xmm4                       \n"
+    "movdqa    %7,%%xmm5                       \n"
     "sub       %1,%2                           \n"
     LABELALIGN
   "1:                                          \n"
     "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
-    BUNDLEALIGN
-    MEMOPREG(movdqu,0x00,0,4,1,xmm7)           //  movdqu  (%0,%4,1),%%xmm7
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)            //  movdqu (%0,%4,1),%%xmm7
     "pavgb     %%xmm7,%%xmm0                   \n"
-    MEMOPREG(movdqu,0x10,0,4,1,xmm7)           //  movdqu  0x10(%0,%4,1),%%xmm7
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)            //  movdqu 0x10(%0,%4,1),%%xmm7
     "pavgb     %%xmm7,%%xmm1                   \n"
-    MEMOPREG(movdqu,0x20,0,4,1,xmm7)           //  movdqu  0x20(%0,%4,1),%%xmm7
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)            //  movdqu 0x20(%0,%4,1),%%xmm7
     "pavgb     %%xmm7,%%xmm2                   \n"
-    MEMOPREG(movdqu,0x30,0,4,1,xmm7)           //  movdqu  0x30(%0,%4,1),%%xmm7
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)            //  movdqu 0x30(%0,%4,1),%%xmm7
     "pavgb     %%xmm7,%%xmm6                   \n"
+
     "lea       " MEMLEA(0x40,0) ",%0           \n"
     "movdqa    %%xmm0,%%xmm7                   \n"
     "shufps    $0x88,%%xmm1,%%xmm0             \n"
@@ -1613,24 +1204,21 @@ void BGRAToUVRow_Unaligned_SSSE3(const uint8* src_bgra0, int src_stride_bgra,
     "psraw     $0x8,%%xmm1                     \n"
     "packsswb  %%xmm1,%%xmm0                   \n"
     "paddb     %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%3                        \n"
     "movlps    %%xmm0," MEMACCESS(1) "         \n"
-    BUNDLEALIGN
     MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
     "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
     "jg        1b                              \n"
   : "+r"(src_bgra0),       // %0
     "+r"(dst_u),           // %1
     "+r"(dst_v),           // %2
     "+rm"(width)           // %3
-  : "r"((intptr_t)(src_stride_bgra)) // %4
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
-#endif
+  : "r"((intptr_t)(src_stride_bgra)), // %4
+    "m"(kBGRAToV),  // %5
+    "m"(kBGRAToU),  // %6
+    "m"(kAddUV128)  // %7
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
   );
 }
 
@@ -1640,43 +1228,6 @@ void ABGRToYRow_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix) {
     "movdqa    %3,%%xmm4                       \n"
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm1                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm4,%%xmm3                   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "phaddw    %%xmm1,%%xmm0                   \n"
-    "phaddw    %%xmm3,%%xmm2                   \n"
-    "psrlw     $0x7,%%xmm0                     \n"
-    "psrlw     $0x7,%%xmm2                     \n"
-    "packuswb  %%xmm2,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "jg        1b                              \n"
-  : "+r"(src_abgr),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(pix)        // %2
-  : "m"(kABGRToY),   // %3
-    "m"(kAddY16)     // %4
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
-  );
-}
-
-void ABGRToYRow_Unaligned_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix) {
-  asm volatile (
-    "movdqa    %4,%%xmm5                       \n"
-    "movdqa    %3,%%xmm4                       \n"
-    LABELALIGN
-  "1:                                          \n"
     "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
     "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
     "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
@@ -1692,19 +1243,16 @@ void ABGRToYRow_Unaligned_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix) {
     "psrlw     $0x7,%%xmm2                     \n"
     "packuswb  %%xmm2,%%xmm0                   \n"
     "paddb     %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
     "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
     "jg        1b                              \n"
   : "+r"(src_abgr),  // %0
     "+r"(dst_y),     // %1
     "+r"(pix)        // %2
   : "m"(kABGRToY),   // %3
     "m"(kAddY16)     // %4
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
 }
 
@@ -1714,43 +1262,6 @@ void RGBAToYRow_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix) {
     "movdqa    %3,%%xmm4                       \n"
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm1                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm4,%%xmm3                   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "phaddw    %%xmm1,%%xmm0                   \n"
-    "phaddw    %%xmm3,%%xmm2                   \n"
-    "psrlw     $0x7,%%xmm0                     \n"
-    "psrlw     $0x7,%%xmm2                     \n"
-    "packuswb  %%xmm2,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "jg        1b                              \n"
-  : "+r"(src_rgba),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(pix)        // %2
-  : "m"(kRGBAToY),   // %3
-    "m"(kAddY16)     // %4
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
-  );
-}
-
-void RGBAToYRow_Unaligned_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix) {
-  asm volatile (
-    "movdqa    %4,%%xmm5                       \n"
-    "movdqa    %3,%%xmm4                       \n"
-    LABELALIGN
-  "1:                                          \n"
     "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
     "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
     "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
@@ -1766,116 +1277,41 @@ void RGBAToYRow_Unaligned_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix) {
     "psrlw     $0x7,%%xmm2                     \n"
     "packuswb  %%xmm2,%%xmm0                   \n"
     "paddb     %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
     "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
     "jg        1b                              \n"
   : "+r"(src_rgba),  // %0
     "+r"(dst_y),     // %1
     "+r"(pix)        // %2
   : "m"(kRGBAToY),   // %3
     "m"(kAddY16)     // %4
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
 }
 
 void ABGRToUVRow_SSSE3(const uint8* src_abgr0, int src_stride_abgr,
                        uint8* dst_u, uint8* dst_v, int width) {
   asm volatile (
-    "movdqa    %0,%%xmm4                       \n"
-    "movdqa    %1,%%xmm3                       \n"
-    "movdqa    %2,%%xmm5                       \n"
-  :
-  : "m"(kABGRToU),         // %0
-    "m"(kABGRToV),         // %1
-    "m"(kAddUV128)         // %2
-  );
-  asm volatile (
-    "sub       %1,%2                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
-    BUNDLEALIGN
-    MEMOPREG(pavgb,0x00,0,4,1,xmm0)            //  pavgb   (%0,%4,1),%%xmm0
-    MEMOPREG(pavgb,0x10,0,4,1,xmm1)            //  pavgb   0x10(%0,%4,1),%%xmm1
-    MEMOPREG(pavgb,0x20,0,4,1,xmm2)            //  pavgb   0x20(%0,%4,1),%%xmm2
-    MEMOPREG(pavgb,0x30,0,4,1,xmm6)            //  pavgb   0x30(%0,%4,1),%%xmm6
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "movdqa    %%xmm0,%%xmm7                   \n"
-    "shufps    $0x88,%%xmm1,%%xmm0             \n"
-    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
-    "pavgb     %%xmm7,%%xmm0                   \n"
-    "movdqa    %%xmm2,%%xmm7                   \n"
-    "shufps    $0x88,%%xmm6,%%xmm2             \n"
-    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
-    "pavgb     %%xmm7,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "movdqa    %%xmm2,%%xmm6                   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm3,%%xmm1                   \n"
-    "pmaddubsw %%xmm3,%%xmm6                   \n"
-    "phaddw    %%xmm2,%%xmm0                   \n"
-    "phaddw    %%xmm6,%%xmm1                   \n"
-    "psraw     $0x8,%%xmm0                     \n"
-    "psraw     $0x8,%%xmm1                     \n"
-    "packsswb  %%xmm1,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%3                        \n"
-    "movlps    %%xmm0," MEMACCESS(1) "         \n"
-    BUNDLEALIGN
-    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "jg        1b                              \n"
-  : "+r"(src_abgr0),       // %0
-    "+r"(dst_u),           // %1
-    "+r"(dst_v),           // %2
-    "+rm"(width)           // %3
-  : "r"((intptr_t)(src_stride_abgr)) // %4
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
-#endif
-  );
-}
-
-void ABGRToUVRow_Unaligned_SSSE3(const uint8* src_abgr0, int src_stride_abgr,
-                                 uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "movdqa    %0,%%xmm4                       \n"
-    "movdqa    %1,%%xmm3                       \n"
-    "movdqa    %2,%%xmm5                       \n"
-  :
-  : "m"(kABGRToU),         // %0
-    "m"(kABGRToV),         // %1
-    "m"(kAddUV128)         // %2
-  );
-  asm volatile (
+    "movdqa    %5,%%xmm3                       \n"
+    "movdqa    %6,%%xmm4                       \n"
+    "movdqa    %7,%%xmm5                       \n"
     "sub       %1,%2                           \n"
     LABELALIGN
   "1:                                          \n"
     "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
-    BUNDLEALIGN
-    MEMOPREG(movdqu,0x00,0,4,1,xmm7)           //  movdqu  (%0,%4,1),%%xmm7
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)            //  movdqu (%0,%4,1),%%xmm7
     "pavgb     %%xmm7,%%xmm0                   \n"
-    MEMOPREG(movdqu,0x10,0,4,1,xmm7)           //  movdqu  0x10(%0,%4,1),%%xmm7
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)            //  movdqu 0x10(%0,%4,1),%%xmm7
     "pavgb     %%xmm7,%%xmm1                   \n"
-    MEMOPREG(movdqu,0x20,0,4,1,xmm7)           //  movdqu  0x20(%0,%4,1),%%xmm7
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)            //  movdqu 0x20(%0,%4,1),%%xmm7
     "pavgb     %%xmm7,%%xmm2                   \n"
-    MEMOPREG(movdqu,0x30,0,4,1,xmm7)           //  movdqu  0x30(%0,%4,1),%%xmm7
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)            //  movdqu 0x30(%0,%4,1),%%xmm7
     "pavgb     %%xmm7,%%xmm6                   \n"
+
     "lea       " MEMLEA(0x40,0) ",%0           \n"
     "movdqa    %%xmm0,%%xmm7                   \n"
     "shufps    $0x88,%%xmm1,%%xmm0             \n"
@@ -1897,121 +1333,46 @@ void ABGRToUVRow_Unaligned_SSSE3(const uint8* src_abgr0, int src_stride_abgr,
     "psraw     $0x8,%%xmm1                     \n"
     "packsswb  %%xmm1,%%xmm0                   \n"
     "paddb     %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%3                        \n"
     "movlps    %%xmm0," MEMACCESS(1) "         \n"
-    BUNDLEALIGN
     MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
     "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
     "jg        1b                              \n"
   : "+r"(src_abgr0),       // %0
     "+r"(dst_u),           // %1
     "+r"(dst_v),           // %2
     "+rm"(width)           // %3
-  : "r"((intptr_t)(src_stride_abgr)) // %4
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
-#endif
+  : "r"((intptr_t)(src_stride_abgr)), // %4
+    "m"(kABGRToV),  // %5
+    "m"(kABGRToU),  // %6
+    "m"(kAddUV128)  // %7
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
   );
 }
 
 void RGBAToUVRow_SSSE3(const uint8* src_rgba0, int src_stride_rgba,
                        uint8* dst_u, uint8* dst_v, int width) {
   asm volatile (
-    "movdqa    %0,%%xmm4                       \n"
-    "movdqa    %1,%%xmm3                       \n"
-    "movdqa    %2,%%xmm5                       \n"
-  :
-  : "m"(kRGBAToU),         // %0
-    "m"(kRGBAToV),         // %1
-    "m"(kAddUV128)         // %2
-  );
-  asm volatile (
-    "sub       %1,%2                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
-    BUNDLEALIGN
-    MEMOPREG(pavgb,0x00,0,4,1,xmm0)            //  pavgb   (%0,%4,1),%%xmm0
-    MEMOPREG(pavgb,0x10,0,4,1,xmm1)            //  pavgb   0x10(%0,%4,1),%%xmm1
-    MEMOPREG(pavgb,0x20,0,4,1,xmm2)            //  pavgb   0x20(%0,%4,1),%%xmm2
-    MEMOPREG(pavgb,0x30,0,4,1,xmm6)            //  pavgb   0x30(%0,%4,1),%%xmm6
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "movdqa    %%xmm0,%%xmm7                   \n"
-    "shufps    $0x88,%%xmm1,%%xmm0             \n"
-    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
-    "pavgb     %%xmm7,%%xmm0                   \n"
-    "movdqa    %%xmm2,%%xmm7                   \n"
-    "shufps    $0x88,%%xmm6,%%xmm2             \n"
-    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
-    "pavgb     %%xmm7,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "movdqa    %%xmm2,%%xmm6                   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm3,%%xmm1                   \n"
-    "pmaddubsw %%xmm3,%%xmm6                   \n"
-    "phaddw    %%xmm2,%%xmm0                   \n"
-    "phaddw    %%xmm6,%%xmm1                   \n"
-    "psraw     $0x8,%%xmm0                     \n"
-    "psraw     $0x8,%%xmm1                     \n"
-    "packsswb  %%xmm1,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%3                        \n"
-    "movlps    %%xmm0," MEMACCESS(1) "         \n"
-    BUNDLEALIGN
-    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "jg        1b                              \n"
-  : "+r"(src_rgba0),       // %0
-    "+r"(dst_u),           // %1
-    "+r"(dst_v),           // %2
-    "+rm"(width)           // %3
-  : "r"((intptr_t)(src_stride_rgba))
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
-#endif
-  );
-}
-
-void RGBAToUVRow_Unaligned_SSSE3(const uint8* src_rgba0, int src_stride_rgba,
-                                 uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "movdqa    %0,%%xmm4                       \n"
-    "movdqa    %1,%%xmm3                       \n"
-    "movdqa    %2,%%xmm5                       \n"
-  :
-  : "m"(kRGBAToU),         // %0
-    "m"(kRGBAToV),         // %1
-    "m"(kAddUV128)         // %2
-  );
-  asm volatile (
+    "movdqa    %5,%%xmm3                       \n"
+    "movdqa    %6,%%xmm4                       \n"
+    "movdqa    %7,%%xmm5                       \n"
     "sub       %1,%2                           \n"
     LABELALIGN
   "1:                                          \n"
     "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
-    BUNDLEALIGN
-    MEMOPREG(movdqu,0x00,0,4,1,xmm7)           //  movdqu  (%0,%4,1),%%xmm7
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)            //  movdqu (%0,%4,1),%%xmm7
     "pavgb     %%xmm7,%%xmm0                   \n"
-    MEMOPREG(movdqu,0x10,0,4,1,xmm7)           //  movdqu  0x10(%0,%4,1),%%xmm7
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)            //  movdqu 0x10(%0,%4,1),%%xmm7
     "pavgb     %%xmm7,%%xmm1                   \n"
-    MEMOPREG(movdqu,0x20,0,4,1,xmm7)           //  movdqu  0x20(%0,%4,1),%%xmm7
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)            //  movdqu 0x20(%0,%4,1),%%xmm7
     "pavgb     %%xmm7,%%xmm2                   \n"
-    MEMOPREG(movdqu,0x30,0,4,1,xmm7)           //  movdqu  0x30(%0,%4,1),%%xmm7
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)            //  movdqu 0x30(%0,%4,1),%%xmm7
     "pavgb     %%xmm7,%%xmm6                   \n"
+
     "lea       " MEMLEA(0x40,0) ",%0           \n"
     "movdqa    %%xmm0,%%xmm7                   \n"
     "shufps    $0x88,%%xmm1,%%xmm0             \n"
@@ -2033,75 +1394,83 @@ void RGBAToUVRow_Unaligned_SSSE3(const uint8* src_rgba0, int src_stride_rgba,
     "psraw     $0x8,%%xmm1                     \n"
     "packsswb  %%xmm1,%%xmm0                   \n"
     "paddb     %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%3                        \n"
     "movlps    %%xmm0," MEMACCESS(1) "         \n"
-    BUNDLEALIGN
     MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
     "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
     "jg        1b                              \n"
   : "+r"(src_rgba0),       // %0
     "+r"(dst_u),           // %1
     "+r"(dst_v),           // %2
     "+rm"(width)           // %3
-  : "r"((intptr_t)(src_stride_rgba)) // %4
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
-#endif
+  : "r"((intptr_t)(src_stride_rgba)), // %4
+    "m"(kRGBAToV),  // %5
+    "m"(kRGBAToU),  // %6
+    "m"(kAddUV128)  // %7
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
   );
 }
-#endif  // HAS_ARGBTOUVROW_SSSE3
 
-#ifdef HAS_I422TOARGBROW_SSSE3
-#define UB 127 /* min(63,(int8)(2.018 * 64)) */
-#define UG -25 /* (int8)(-0.391 * 64 - 0.5) */
-#define UR 0
-
-#define VB 0
-#define VG -52 /* (int8)(-0.813 * 64 - 0.5) */
-#define VR 102 /* (int8)(1.596 * 64 + 0.5) */
-
-// Bias
-#define BB UB * 128 + VB * 128
-#define BG UG * 128 + VG * 128
-#define BR UR * 128 + VR * 128
-
-#define YG 74 /* (int8)(1.164 * 64 + 0.5) */
-
-struct {
-  vec8 kUVToB;  // 0
-  vec8 kUVToG;  // 16
-  vec8 kUVToR;  // 32
-  vec16 kUVBiasB;  // 48
-  vec16 kUVBiasG;  // 64
-  vec16 kUVBiasR;  // 80
-  vec16 kYSub16;  // 96
-  vec16 kYToRgb;  // 112
-  vec8 kVUToB;  // 128
-  vec8 kVUToG;  // 144
-  vec8 kVUToR;  // 160
-} static SIMD_ALIGNED(kYuvConstants) = {
-  { UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB },
-  { UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG },
-  { UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR },
-  { BB, BB, BB, BB, BB, BB, BB, BB },
-  { BG, BG, BG, BG, BG, BG, BG, BG },
-  { BR, BR, BR, BR, BR, BR, BR, BR },
-  { 16, 16, 16, 16, 16, 16, 16, 16 },
-  { YG, YG, YG, YG, YG, YG, YG, YG },
-  { VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB },
-  { VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG },
-  { VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR }
+#if defined(HAS_I422TOARGBROW_SSSE3) || defined(HAS_I422TOARGBROW_AVX2)
+
+// YUV to RGB conversion constants.
+// Y contribution to R,G,B.  Scale and bias.
+#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */
+#define YGB 1160 /* 1.164 * 64 * 16 - adjusted for even error distribution */
+
+// U and V contributions to R,G,B.
+#define UB -128 /* -min(128, round(2.018 * 64)) */
+#define UG 25 /* -round(-0.391 * 64) */
+#define VG 52 /* -round(-0.813 * 64) */
+#define VR -102 /* -round(1.596 * 64) */
+
+// Bias values to subtract 16 from Y and 128 from U and V.
+#define BB (UB * 128            - YGB)
+#define BG (UG * 128 + VG * 128 - YGB)
+#define BR            (VR * 128 - YGB)
+
+struct YuvConstants {
+  lvec8 kUVToB;     // 0
+  lvec8 kUVToG;     // 32
+  lvec8 kUVToR;     // 64
+  lvec16 kUVBiasB;  // 96
+  lvec16 kUVBiasG;  // 128
+  lvec16 kUVBiasR;  // 160
+  lvec16 kYToRgb;   // 192
+};
+
+// BT601 constants for YUV to RGB.
+static YuvConstants SIMD_ALIGNED(kYuvConstants) = {
+  { UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0,
+    UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0 },
+  { UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG,
+    UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG },
+  { 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR,
+    0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR },
+  { BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB },
+  { BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG },
+  { BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR },
+  { YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG }
 };
 
+// BT601 constants for NV21 where chroma plane is VU instead of UV.
+static YuvConstants SIMD_ALIGNED(kYvuConstants) = {
+  { 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB,
+    0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB },
+  { VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG,
+    VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG },
+  { VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0,
+    VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0 },
+  { BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB },
+  { BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG },
+  { BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR },
+  { YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG }
+};
 
 // Read 8 UV from 411
 #define READYUV444                                                             \
     "movq       " MEMACCESS([u_buf]) ",%%xmm0                   \n"            \
-    BUNDLEALIGN                                                                \
     MEMOPREG(movq, 0x00, [u_buf], [v_buf], 1, xmm1)                            \
     "lea        " MEMLEA(0x8, [u_buf]) ",%[u_buf]               \n"            \
     "punpcklbw  %%xmm1,%%xmm0                                   \n"
@@ -2109,7 +1478,6 @@ struct {
 // Read 4 UV from 422, upsample to 8 UV
 #define READYUV422                                                             \
     "movd       " MEMACCESS([u_buf]) ",%%xmm0                   \n"            \
-    BUNDLEALIGN                                                                \
     MEMOPREG(movd, 0x00, [u_buf], [v_buf], 1, xmm1)                            \
     "lea        " MEMLEA(0x4, [u_buf]) ",%[u_buf]               \n"            \
     "punpcklbw  %%xmm1,%%xmm0                                   \n"            \
@@ -2118,7 +1486,6 @@ struct {
 // Read 2 UV from 411, upsample to 8 UV
 #define READYUV411                                                             \
     "movd       " MEMACCESS([u_buf]) ",%%xmm0                   \n"            \
-    BUNDLEALIGN                                                                \
     MEMOPREG(movd, 0x00, [u_buf], [v_buf], 1, xmm1)                            \
     "lea        " MEMLEA(0x2, [u_buf]) ",%[u_buf]               \n"            \
     "punpcklbw  %%xmm1,%%xmm0                                   \n"            \
@@ -2132,20 +1499,23 @@ struct {
     "punpcklwd  %%xmm0,%%xmm0                                   \n"
 
 // Convert 8 pixels: 8 UV and 8 Y
-#define YUVTORGB                                                               \
+#define YUVTORGB(YuvConstants)                                                 \
     "movdqa     %%xmm0,%%xmm1                                   \n"            \
     "movdqa     %%xmm0,%%xmm2                                   \n"            \
-    "pmaddubsw  " MEMACCESS([kYuvConstants]) ",%%xmm0           \n"            \
-    "pmaddubsw  " MEMACCESS2(16, [kYuvConstants]) ",%%xmm1      \n"            \
-    "pmaddubsw  " MEMACCESS2(32, [kYuvConstants]) ",%%xmm2      \n"            \
-    "psubw      " MEMACCESS2(48, [kYuvConstants]) ",%%xmm0      \n"            \
-    "psubw      " MEMACCESS2(64, [kYuvConstants]) ",%%xmm1      \n"            \
-    "psubw      " MEMACCESS2(80, [kYuvConstants]) ",%%xmm2      \n"            \
+    "movdqa     %%xmm0,%%xmm3                                   \n"            \
+    "movdqa     " MEMACCESS2(96, [YuvConstants]) ",%%xmm0       \n"            \
+    "pmaddubsw  " MEMACCESS([YuvConstants]) ",%%xmm1            \n"            \
+    "psubw      %%xmm1,%%xmm0                                   \n"            \
+    "movdqa     " MEMACCESS2(128, [YuvConstants]) ",%%xmm1      \n"            \
+    "pmaddubsw  " MEMACCESS2(32, [YuvConstants]) ",%%xmm2       \n"            \
+    "psubw      %%xmm2,%%xmm1                                   \n"            \
+    "movdqa     " MEMACCESS2(160, [YuvConstants]) ",%%xmm2      \n"            \
+    "pmaddubsw  " MEMACCESS2(64, [YuvConstants]) ",%%xmm3       \n"            \
+    "psubw      %%xmm3,%%xmm2                                   \n"            \
     "movq       " MEMACCESS([y_buf]) ",%%xmm3                   \n"            \
     "lea        " MEMLEA(0x8, [y_buf]) ",%[y_buf]               \n"            \
-    "punpcklbw  %%xmm4,%%xmm3                                   \n"            \
-    "psubsw     " MEMACCESS2(96, [kYuvConstants]) ",%%xmm3      \n"            \
-    "pmullw     " MEMACCESS2(112, [kYuvConstants]) ",%%xmm3     \n"            \
+    "punpcklbw  %%xmm3,%%xmm3                                   \n"            \
+    "pmulhuw    " MEMACCESS2(192, [YuvConstants]) ",%%xmm3      \n"            \
     "paddsw     %%xmm3,%%xmm0                                   \n"            \
     "paddsw     %%xmm3,%%xmm1                                   \n"            \
     "paddsw     %%xmm3,%%xmm2                                   \n"            \
@@ -2156,30 +1526,51 @@ struct {
     "packuswb   %%xmm1,%%xmm1                                   \n"            \
     "packuswb   %%xmm2,%%xmm2                                   \n"
 
-// Convert 8 pixels: 8 VU and 8 Y
-#define YVUTORGB                                                               \
-    "movdqa     %%xmm0,%%xmm1                                   \n"            \
-    "movdqa     %%xmm0,%%xmm2                                   \n"            \
-    "pmaddubsw  " MEMACCESS2(128, [kYuvConstants]) ",%%xmm0     \n"            \
-    "pmaddubsw  " MEMACCESS2(144, [kYuvConstants]) ",%%xmm1     \n"            \
-    "pmaddubsw  " MEMACCESS2(160, [kYuvConstants]) ",%%xmm2     \n"            \
-    "psubw      " MEMACCESS2(48, [kYuvConstants]) ",%%xmm0      \n"            \
-    "psubw      " MEMACCESS2(64, [kYuvConstants]) ",%%xmm1      \n"            \
-    "psubw      " MEMACCESS2(80, [kYuvConstants]) ",%%xmm2      \n"            \
-    "movq       " MEMACCESS([y_buf]) ",%%xmm3                   \n"            \
-    "lea        " MEMLEA(0x8, [y_buf]) ",%[y_buf]               \n"            \
-    "punpcklbw  %%xmm4,%%xmm3                                   \n"            \
-    "psubsw     " MEMACCESS2(96, [kYuvConstants]) ",%%xmm3      \n"            \
-    "pmullw     " MEMACCESS2(112, [kYuvConstants]) ",%%xmm3     \n"            \
-    "paddsw     %%xmm3,%%xmm0                                   \n"            \
-    "paddsw     %%xmm3,%%xmm1                                   \n"            \
-    "paddsw     %%xmm3,%%xmm2                                   \n"            \
-    "psraw      $0x6,%%xmm0                                     \n"            \
-    "psraw      $0x6,%%xmm1                                     \n"            \
-    "psraw      $0x6,%%xmm2                                     \n"            \
-    "packuswb   %%xmm0,%%xmm0                                   \n"            \
-    "packuswb   %%xmm1,%%xmm1                                   \n"            \
-    "packuswb   %%xmm2,%%xmm2                                   \n"
+// Store 8 ARGB values. Assumes XMM5 is zero.
+#define STOREARGB                                                              \
+    "punpcklbw  %%xmm1,%%xmm0                                    \n"           \
+    "punpcklbw  %%xmm5,%%xmm2                                    \n"           \
+    "movdqa     %%xmm0,%%xmm1                                    \n"           \
+    "punpcklwd  %%xmm2,%%xmm0                                    \n"           \
+    "punpckhwd  %%xmm2,%%xmm1                                    \n"           \
+    "movdqu     %%xmm0," MEMACCESS([dst_argb]) "                 \n"           \
+    "movdqu     %%xmm1," MEMACCESS2(0x10,[dst_argb]) "           \n"           \
+    "lea        " MEMLEA(0x20,[dst_argb]) ",%[dst_argb]          \n"
+
+// Store 8 BGRA values. Assumes XMM5 is zero.
+#define STOREBGRA                                                              \
+    "pcmpeqb   %%xmm5,%%xmm5                                     \n"           \
+    "punpcklbw %%xmm0,%%xmm1                                     \n"           \
+    "punpcklbw %%xmm2,%%xmm5                                     \n"           \
+    "movdqa    %%xmm5,%%xmm0                                     \n"           \
+    "punpcklwd %%xmm1,%%xmm5                                     \n"           \
+    "punpckhwd %%xmm1,%%xmm0                                     \n"           \
+    "movdqu    %%xmm5," MEMACCESS([dst_bgra]) "                  \n"           \
+    "movdqu    %%xmm0," MEMACCESS2(0x10,[dst_bgra]) "            \n"           \
+    "lea       " MEMLEA(0x20,[dst_bgra]) ",%[dst_bgra]           \n"
+
+// Store 8 ABGR values. Assumes XMM5 is zero.
+#define STOREABGR                                                              \
+    "punpcklbw %%xmm1,%%xmm2                                     \n"           \
+    "punpcklbw %%xmm5,%%xmm0                                     \n"           \
+    "movdqa    %%xmm2,%%xmm1                                     \n"           \
+    "punpcklwd %%xmm0,%%xmm2                                     \n"           \
+    "punpckhwd %%xmm0,%%xmm1                                     \n"           \
+    "movdqu    %%xmm2," MEMACCESS([dst_abgr]) "                  \n"           \
+    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_abgr]) "            \n"           \
+    "lea       " MEMLEA(0x20,[dst_abgr]) ",%[dst_abgr]           \n"
+
+// Store 8 RGBA values. Assumes XMM5 is zero.
+#define STORERGBA                                                              \
+    "pcmpeqb   %%xmm5,%%xmm5                                     \n"           \
+    "punpcklbw %%xmm2,%%xmm1                                     \n"           \
+    "punpcklbw %%xmm0,%%xmm5                                     \n"           \
+    "movdqa    %%xmm5,%%xmm0                                     \n"           \
+    "punpcklwd %%xmm1,%%xmm5                                     \n"           \
+    "punpckhwd %%xmm1,%%xmm0                                     \n"           \
+    "movdqu    %%xmm5," MEMACCESS([dst_rgba]) "                  \n"           \
+    "movdqu    %%xmm0," MEMACCESS2(0x10,[dst_rgba]) "            \n"           \
+    "lea       " MEMLEA(0x20,[dst_rgba]) ",%[dst_rgba]           \n"
 
 void OMITFP I444ToARGBRow_SSSE3(const uint8* y_buf,
                                 const uint8* u_buf,
@@ -2189,19 +1580,11 @@ void OMITFP I444ToARGBRow_SSSE3(const uint8* y_buf,
   asm volatile (
     "sub       %[u_buf],%[v_buf]               \n"
     "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "pxor      %%xmm4,%%xmm4                   \n"
     LABELALIGN
   "1:                                          \n"
     READYUV444
-    YUVTORGB
-    "punpcklbw %%xmm1,%%xmm0                   \n"
-    "punpcklbw %%xmm5,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpcklwd %%xmm2,%%xmm0                   \n"
-    "punpckhwd %%xmm2,%%xmm1                   \n"
-    "movdqa    %%xmm0," MEMACCESS([dst_argb]) "         \n"
-    "movdqa    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "   \n"
-    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb]  \n"
+    YUVTORGB(kYuvConstants)
+    STOREARGB
     "sub       $0x8,%[width]                   \n"
     "jg        1b                              \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
@@ -2210,41 +1593,25 @@ void OMITFP I444ToARGBRow_SSSE3(const uint8* y_buf,
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
   : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
   );
 }
 
+// TODO(fbarchard): Consider putting masks into constants.
 void OMITFP I422ToRGB24Row_SSSE3(const uint8* y_buf,
                                  const uint8* u_buf,
                                  const uint8* v_buf,
                                  uint8* dst_rgb24,
                                  int width) {
-// fpic 32 bit gcc 4.2 on OSX runs out of GPR regs.
-#if defined(__i386__)
-  asm volatile (
-    "movdqa    %[kShuffleMaskARGBToRGB24_0],%%xmm5 \n"
-    "movdqa    %[kShuffleMaskARGBToRGB24],%%xmm6   \n"
-  :: [kShuffleMaskARGBToRGB24_0]"m"(kShuffleMaskARGBToRGB24_0),
-    [kShuffleMaskARGBToRGB24]"m"(kShuffleMaskARGBToRGB24));
-#endif
-
   asm volatile (
-#if !defined(__i386__)
     "movdqa    %[kShuffleMaskARGBToRGB24_0],%%xmm5 \n"
     "movdqa    %[kShuffleMaskARGBToRGB24],%%xmm6   \n"
-#endif
     "sub       %[u_buf],%[v_buf]               \n"
-    "pxor      %%xmm4,%%xmm4                   \n"
     LABELALIGN
   "1:                                          \n"
     READYUV422
-    YUVTORGB
+    YUVTORGB(kYuvConstants)
     "punpcklbw %%xmm1,%%xmm0                   \n"
     "punpcklbw %%xmm2,%%xmm2                   \n"
     "movdqa    %%xmm0,%%xmm1                   \n"
@@ -2256,25 +1623,23 @@ void OMITFP I422ToRGB24Row_SSSE3(const uint8* y_buf,
     "movq      %%xmm0," MEMACCESS([dst_rgb24]) "\n"
     "movdqu    %%xmm1," MEMACCESS2(0x8,[dst_rgb24]) "\n"
     "lea       " MEMLEA(0x18,[dst_rgb24]) ",%[dst_rgb24] \n"
-    "sub       $0x8,%[width]                   \n"
+    "subl      $0x8,%[width]                   \n"
     "jg        1b                              \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
     [dst_rgb24]"+r"(dst_rgb24),  // %[dst_rgb24]
+// TODO(fbarchard): Make width a register for 32 bit.
+#if defined(__i386__) && defined(__pic__)
+    [width]"+m"(width)     // %[width]
+#else
     [width]"+rm"(width)    // %[width]
-  : [kYuvConstants]"r"(&kYuvConstants.kUVToB)
-#if !defined(__i386__)
-    , [kShuffleMaskARGBToRGB24_0]"m"(kShuffleMaskARGBToRGB24_0),
-    [kShuffleMaskARGBToRGB24]"m"(kShuffleMaskARGBToRGB24)
-#endif
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
 #endif
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB),
+    [kShuffleMaskARGBToRGB24_0]"m"(kShuffleMaskARGBToRGB24_0),
+    [kShuffleMaskARGBToRGB24]"m"(kShuffleMaskARGBToRGB24)
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5", "xmm6"
   );
 }
 
@@ -2283,26 +1648,14 @@ void OMITFP I422ToRAWRow_SSSE3(const uint8* y_buf,
                                const uint8* v_buf,
                                uint8* dst_raw,
                                int width) {
-// fpic 32 bit gcc 4.2 on OSX runs out of GPR regs.
-#if defined(__i386__)
-  asm volatile (
-    "movdqa    %[kShuffleMaskARGBToRAW_0],%%xmm5 \n"
-    "movdqa    %[kShuffleMaskARGBToRAW],%%xmm6   \n"
-  :: [kShuffleMaskARGBToRAW_0]"m"(kShuffleMaskARGBToRAW_0),
-    [kShuffleMaskARGBToRAW]"m"(kShuffleMaskARGBToRAW));
-#endif
-
   asm volatile (
-#if !defined(__i386__)
     "movdqa    %[kShuffleMaskARGBToRAW_0],%%xmm5 \n"
     "movdqa    %[kShuffleMaskARGBToRAW],%%xmm6   \n"
-#endif
     "sub       %[u_buf],%[v_buf]               \n"
-    "pxor      %%xmm4,%%xmm4                   \n"
     LABELALIGN
   "1:                                          \n"
     READYUV422
-    YUVTORGB
+    YUVTORGB(kYuvConstants)
     "punpcklbw %%xmm1,%%xmm0                   \n"
     "punpcklbw %%xmm2,%%xmm2                   \n"
     "movdqa    %%xmm0,%%xmm1                   \n"
@@ -2314,25 +1667,23 @@ void OMITFP I422ToRAWRow_SSSE3(const uint8* y_buf,
     "movq      %%xmm0," MEMACCESS([dst_raw]) " \n"
     "movdqu    %%xmm1," MEMACCESS2(0x8,[dst_raw]) "\n"
     "lea       " MEMLEA(0x18,[dst_raw]) ",%[dst_raw] \n"
-    "sub       $0x8,%[width]                   \n"
+    "subl      $0x8,%[width]                   \n"
     "jg        1b                              \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
     [dst_raw]"+r"(dst_raw),  // %[dst_raw]
+// TODO(fbarchard): Make width a register for 32 bit.
+#if defined(__i386__) && defined(__pic__)
+    [width]"+m"(width)    // %[width]
+#else
     [width]"+rm"(width)    // %[width]
-  : [kYuvConstants]"r"(&kYuvConstants.kUVToB)
-#if !defined(__i386__)
-    , [kShuffleMaskARGBToRAW_0]"m"(kShuffleMaskARGBToRAW_0),
-    [kShuffleMaskARGBToRAW]"m"(kShuffleMaskARGBToRAW)
-#endif
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
 #endif
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB),
+    [kShuffleMaskARGBToRAW_0]"m"(kShuffleMaskARGBToRAW_0),
+    [kShuffleMaskARGBToRAW]"m"(kShuffleMaskARGBToRAW)
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5", "xmm6"
   );
 }
 
@@ -2344,19 +1695,11 @@ void OMITFP I422ToARGBRow_SSSE3(const uint8* y_buf,
   asm volatile (
     "sub       %[u_buf],%[v_buf]               \n"
     "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "pxor      %%xmm4,%%xmm4                   \n"
     LABELALIGN
   "1:                                          \n"
     READYUV422
-    YUVTORGB
-    "punpcklbw %%xmm1,%%xmm0                   \n"
-    "punpcklbw %%xmm5,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpcklwd %%xmm2,%%xmm0                   \n"
-    "punpckhwd %%xmm2,%%xmm1                   \n"
-    "movdqa    %%xmm0," MEMACCESS([dst_argb]) "\n"
-    "movdqa    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
-    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    YUVTORGB(kYuvConstants)
+    STOREARGB
     "sub       $0x8,%[width]                   \n"
     "jg        1b                              \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
@@ -2365,13 +1708,8 @@ void OMITFP I422ToARGBRow_SSSE3(const uint8* y_buf,
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
   : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
   );
 }
 
@@ -2383,19 +1721,11 @@ void OMITFP I411ToARGBRow_SSSE3(const uint8* y_buf,
   asm volatile (
     "sub       %[u_buf],%[v_buf]               \n"
     "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "pxor      %%xmm4,%%xmm4                   \n"
     LABELALIGN
   "1:                                          \n"
     READYUV411
-    YUVTORGB
-    "punpcklbw %%xmm1,%%xmm0                   \n"
-    "punpcklbw %%xmm5,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpcklwd %%xmm2,%%xmm0                   \n"
-    "punpckhwd %%xmm2,%%xmm1                   \n"
-    "movdqa    %%xmm0," MEMACCESS([dst_argb]) "\n"
-    "movdqa    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
-    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    YUVTORGB(kYuvConstants)
+    STOREARGB
     "sub       $0x8,%[width]                   \n"
     "jg        1b                              \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
@@ -2404,13 +1734,8 @@ void OMITFP I411ToARGBRow_SSSE3(const uint8* y_buf,
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
   : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
   );
 }
 
@@ -2420,19 +1745,11 @@ void OMITFP NV12ToARGBRow_SSSE3(const uint8* y_buf,
                                 int width) {
   asm volatile (
     "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "pxor      %%xmm4,%%xmm4                   \n"
     LABELALIGN
   "1:                                          \n"
     READNV12
-    YUVTORGB
-    "punpcklbw %%xmm1,%%xmm0                   \n"
-    "punpcklbw %%xmm5,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpcklwd %%xmm2,%%xmm0                   \n"
-    "punpckhwd %%xmm2,%%xmm1                   \n"
-    "movdqa    %%xmm0," MEMACCESS([dst_argb]) "\n"
-    "movdqa    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
-    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    YUVTORGB(kYuvConstants)
+    STOREARGB
     "sub       $0x8,%[width]                   \n"
     "jg        1b                              \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
@@ -2440,11 +1757,8 @@ void OMITFP NV12ToARGBRow_SSSE3(const uint8* y_buf,
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
   : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
-  : "memory", "cc"
   // Does not use r14.
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
   );
 }
 
@@ -2454,216 +1768,20 @@ void OMITFP NV21ToARGBRow_SSSE3(const uint8* y_buf,
                                 int width) {
   asm volatile (
     "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "pxor      %%xmm4,%%xmm4                   \n"
-    LABELALIGN
-  "1:                                          \n"
-    READNV12
-    YVUTORGB
-    "punpcklbw %%xmm1,%%xmm0                   \n"
-    "punpcklbw %%xmm5,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpcklwd %%xmm2,%%xmm0                   \n"
-    "punpckhwd %%xmm2,%%xmm1                   \n"
-    "movdqa    %%xmm0," MEMACCESS([dst_argb]) "\n"
-    "movdqa    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
-    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
-    "sub       $0x8,%[width]                   \n"
-    "jg        1b                              \n"
-  : [y_buf]"+r"(y_buf),    // %[y_buf]
-    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
-    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
-    [width]"+rm"(width)    // %[width]
-  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
-  : "memory", "cc"
-  // Does not use r14.
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
-  );
-}
-
-void OMITFP I444ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
-                                          const uint8* u_buf,
-                                          const uint8* v_buf,
-                                          uint8* dst_argb,
-                                          int width) {
-  asm volatile (
-    "sub       %[u_buf],%[v_buf]               \n"
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "pxor      %%xmm4,%%xmm4                   \n"
-    LABELALIGN
-  "1:                                          \n"
-    READYUV444
-    YUVTORGB
-    "punpcklbw %%xmm1,%%xmm0                   \n"
-    "punpcklbw %%xmm5,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpcklwd %%xmm2,%%xmm0                   \n"
-    "punpckhwd %%xmm2,%%xmm1                   \n"
-    "movdqu    %%xmm0," MEMACCESS([dst_argb]) "\n"
-    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
-    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
-    "sub       $0x8,%[width]                   \n"
-    "jg        1b                              \n"
-  : [y_buf]"+r"(y_buf),    // %[y_buf]
-    [u_buf]"+r"(u_buf),    // %[u_buf]
-    [v_buf]"+r"(v_buf),    // %[v_buf]
-    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
-    [width]"+rm"(width)    // %[width]
-  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
-  );
-}
-
-void OMITFP I422ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
-                                          const uint8* u_buf,
-                                          const uint8* v_buf,
-                                          uint8* dst_argb,
-                                          int width) {
-  asm volatile (
-    "sub       %[u_buf],%[v_buf]               \n"
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "pxor      %%xmm4,%%xmm4                   \n"
-    LABELALIGN
-  "1:                                          \n"
-    READYUV422
-    YUVTORGB
-    "punpcklbw %%xmm1,%%xmm0                   \n"
-    "punpcklbw %%xmm5,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpcklwd %%xmm2,%%xmm0                   \n"
-    "punpckhwd %%xmm2,%%xmm1                   \n"
-    "movdqu    %%xmm0," MEMACCESS([dst_argb]) "\n"
-    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
-    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
-    "sub       $0x8,%[width]                   \n"
-    "jg        1b                              \n"
-  : [y_buf]"+r"(y_buf),    // %[y_buf]
-    [u_buf]"+r"(u_buf),    // %[u_buf]
-    [v_buf]"+r"(v_buf),    // %[v_buf]
-    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
-    [width]"+rm"(width)    // %[width]
-  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
-  );
-}
-
-void OMITFP I411ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
-                                          const uint8* u_buf,
-                                          const uint8* v_buf,
-                                          uint8* dst_argb,
-                                          int width) {
-  asm volatile (
-    "sub       %[u_buf],%[v_buf]               \n"
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "pxor      %%xmm4,%%xmm4                   \n"
-    LABELALIGN
-  "1:                                          \n"
-    READYUV411
-    YUVTORGB
-    "punpcklbw %%xmm1,%%xmm0                   \n"
-    "punpcklbw %%xmm5,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpcklwd %%xmm2,%%xmm0                   \n"
-    "punpckhwd %%xmm2,%%xmm1                   \n"
-    "movdqu    %%xmm0," MEMACCESS([dst_argb]) "\n"
-    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
-    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
-    "sub       $0x8,%[width]                   \n"
-    "jg        1b                              \n"
-  : [y_buf]"+r"(y_buf),    // %[y_buf]
-    [u_buf]"+r"(u_buf),    // %[u_buf]
-    [v_buf]"+r"(v_buf),    // %[v_buf]
-    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
-    [width]"+rm"(width)    // %[width]
-  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
-  );
-}
-
-void OMITFP NV12ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
-                                          const uint8* uv_buf,
-                                          uint8* dst_argb,
-                                          int width) {
-  asm volatile (
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "pxor      %%xmm4,%%xmm4                   \n"
     LABELALIGN
   "1:                                          \n"
     READNV12
-    YUVTORGB
-    "punpcklbw %%xmm1,%%xmm0                   \n"
-    "punpcklbw %%xmm5,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpcklwd %%xmm2,%%xmm0                   \n"
-    "punpckhwd %%xmm2,%%xmm1                   \n"
-    "movdqu    %%xmm0," MEMACCESS([dst_argb]) "\n"
-    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
-    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    YUVTORGB(kYuvConstants)
+    STOREARGB
     "sub       $0x8,%[width]                   \n"
     "jg        1b                              \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [uv_buf]"+r"(uv_buf),    // %[uv_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
-  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
-  : "memory", "cc"
-  // Does not use r14.
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
-  );
-}
-
-void OMITFP NV21ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
-                                          const uint8* uv_buf,
-                                          uint8* dst_argb,
-                                          int width) {
-  asm volatile (
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "pxor      %%xmm4,%%xmm4                   \n"
-    LABELALIGN
-  "1:                                          \n"
-    READNV12
-    YVUTORGB
-    "punpcklbw %%xmm1,%%xmm0                   \n"
-    "punpcklbw %%xmm5,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpcklwd %%xmm2,%%xmm0                   \n"
-    "punpckhwd %%xmm2,%%xmm1                   \n"
-    "movdqu    %%xmm0," MEMACCESS([dst_argb]) "\n"
-    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
-    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
-    "sub       $0x8,%[width]                   \n"
-    "jg        1b                              \n"
-  : [y_buf]"+r"(y_buf),    // %[y_buf]
-    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
-    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
-    [width]"+rm"(width)    // %[width]
-  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
-  : "memory", "cc"
+  : [kYuvConstants]"r"(&kYvuConstants.kUVToB) // %[kYuvConstants]
   // Does not use r14.
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
   );
 }
 
@@ -2675,20 +1793,11 @@ void OMITFP I422ToBGRARow_SSSE3(const uint8* y_buf,
   asm volatile (
     "sub       %[u_buf],%[v_buf]               \n"
     "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "pxor      %%xmm4,%%xmm4                   \n"
     LABELALIGN
   "1:                                          \n"
     READYUV422
-    YUVTORGB
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "punpcklbw %%xmm0,%%xmm1                   \n"
-    "punpcklbw %%xmm2,%%xmm5                   \n"
-    "movdqa    %%xmm5,%%xmm0                   \n"
-    "punpcklwd %%xmm1,%%xmm5                   \n"
-    "punpckhwd %%xmm1,%%xmm0                   \n"
-    "movdqa    %%xmm5," MEMACCESS([dst_bgra]) "\n"
-    "movdqa    %%xmm0," MEMACCESS2(0x10,[dst_bgra]) "\n"
-    "lea       " MEMLEA(0x20,[dst_bgra]) ",%[dst_bgra] \n"
+    YUVTORGB(kYuvConstants)
+    STOREBGRA
     "sub       $0x8,%[width]                   \n"
     "jg        1b                              \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
@@ -2697,13 +1806,8 @@ void OMITFP I422ToBGRARow_SSSE3(const uint8* y_buf,
     [dst_bgra]"+r"(dst_bgra),  // %[dst_bgra]
     [width]"+rm"(width)    // %[width]
   : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
   );
 }
 
@@ -2715,19 +1819,11 @@ void OMITFP I422ToABGRRow_SSSE3(const uint8* y_buf,
   asm volatile (
     "sub       %[u_buf],%[v_buf]               \n"
     "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "pxor      %%xmm4,%%xmm4                   \n"
     LABELALIGN
   "1:                                          \n"
     READYUV422
-    YUVTORGB
-    "punpcklbw %%xmm1,%%xmm2                   \n"
-    "punpcklbw %%xmm5,%%xmm0                   \n"
-    "movdqa    %%xmm2,%%xmm1                   \n"
-    "punpcklwd %%xmm0,%%xmm2                   \n"
-    "punpckhwd %%xmm0,%%xmm1                   \n"
-    "movdqa    %%xmm2," MEMACCESS([dst_abgr]) "\n"
-    "movdqa    %%xmm1," MEMACCESS2(0x10,[dst_abgr]) "\n"
-    "lea       " MEMLEA(0x20,[dst_abgr]) ",%[dst_abgr] \n"
+    YUVTORGB(kYuvConstants)
+    STOREABGR
     "sub       $0x8,%[width]                   \n"
     "jg        1b                              \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
@@ -2736,13 +1832,8 @@ void OMITFP I422ToABGRRow_SSSE3(const uint8* y_buf,
     [dst_abgr]"+r"(dst_abgr),  // %[dst_abgr]
     [width]"+rm"(width)    // %[width]
   : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
   );
 }
 
@@ -2754,20 +1845,11 @@ void OMITFP I422ToRGBARow_SSSE3(const uint8* y_buf,
   asm volatile (
     "sub       %[u_buf],%[v_buf]               \n"
     "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "pxor      %%xmm4,%%xmm4                   \n"
     LABELALIGN
   "1:                                          \n"
     READYUV422
-    YUVTORGB
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "punpcklbw %%xmm2,%%xmm1                   \n"
-    "punpcklbw %%xmm0,%%xmm5                   \n"
-    "movdqa    %%xmm5,%%xmm0                   \n"
-    "punpcklwd %%xmm1,%%xmm5                   \n"
-    "punpckhwd %%xmm1,%%xmm0                   \n"
-    "movdqa    %%xmm5," MEMACCESS([dst_rgba]) "\n"
-    "movdqa    %%xmm0," MEMACCESS2(0x10,[dst_rgba]) "\n"
-    "lea       " MEMLEA(0x20,[dst_rgba]) ",%[dst_rgba] \n"
+    YUVTORGB(kYuvConstants)
+    STORERGBA
     "sub       $0x8,%[width]                   \n"
     "jg        1b                              \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
@@ -2776,159 +1858,233 @@ void OMITFP I422ToRGBARow_SSSE3(const uint8* y_buf,
     [dst_rgba]"+r"(dst_rgba),  // %[dst_rgba]
     [width]"+rm"(width)    // %[width]
   : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
   );
 }
 
-void OMITFP I422ToBGRARow_Unaligned_SSSE3(const uint8* y_buf,
-                                          const uint8* u_buf,
-                                          const uint8* v_buf,
-                                          uint8* dst_bgra,
-                                          int width) {
+#endif  // HAS_I422TOARGBROW_SSSE3
+
+// Read 8 UV from 422, upsample to 16 UV.
+#define READYUV422_AVX2                                                        \
+    "vmovq       " MEMACCESS([u_buf]) ",%%xmm0                      \n"        \
+    MEMOPREG(vmovq, 0x00, [u_buf], [v_buf], 1, xmm1)                           \
+    "lea        " MEMLEA(0x8, [u_buf]) ",%[u_buf]                   \n"        \
+    "vpunpcklbw %%ymm1,%%ymm0,%%ymm0                                \n"        \
+    "vpermq     $0xd8,%%ymm0,%%ymm0                                 \n"        \
+    "vpunpcklwd %%ymm0,%%ymm0,%%ymm0                                \n"
+
+// Convert 16 pixels: 16 UV and 16 Y.
+#define YUVTORGB_AVX2(YuvConstants)                                            \
+    "vpmaddubsw  " MEMACCESS2(64, [YuvConstants]) ",%%ymm0,%%ymm2   \n"        \
+    "vpmaddubsw  " MEMACCESS2(32, [YuvConstants]) ",%%ymm0,%%ymm1   \n"        \
+    "vpmaddubsw  " MEMACCESS([YuvConstants]) ",%%ymm0,%%ymm0        \n"        \
+    "vmovdqu     " MEMACCESS2(160, [YuvConstants]) ",%%ymm3         \n"        \
+    "vpsubw      %%ymm2,%%ymm3,%%ymm2                               \n"        \
+    "vmovdqu     " MEMACCESS2(128, [YuvConstants]) ",%%ymm2         \n"        \
+    "vpsubw      %%ymm1,%%ymm2,%%ymm1                               \n"        \
+    "vmovdqu     " MEMACCESS2(96, [YuvConstants]) ",%%ymm1          \n"        \
+    "vpsubw      %%ymm0,%%ymm1,%%ymm0                               \n"        \
+    "vmovdqu     " MEMACCESS([y_buf]) ",%%xmm3                      \n"        \
+    "lea         " MEMLEA(0x10, [y_buf]) ",%[y_buf]                 \n"        \
+    "vpermq      $0xd8,%%ymm3,%%ymm3                                \n"        \
+    "vpunpcklbw  %%ymm3,%%ymm3,%%ymm3                               \n"        \
+    "vpmulhuw    " MEMACCESS2(192, [YuvConstants]) ",%%ymm3,%%ymm3  \n"        \
+    "vpaddsw     %%ymm3,%%ymm0,%%ymm0           \n"                            \
+    "vpaddsw     %%ymm3,%%ymm1,%%ymm1           \n"                            \
+    "vpaddsw     %%ymm3,%%ymm2,%%ymm2           \n"                            \
+    "vpsraw      $0x6,%%ymm0,%%ymm0             \n"                            \
+    "vpsraw      $0x6,%%ymm1,%%ymm1             \n"                            \
+    "vpsraw      $0x6,%%ymm2,%%ymm2             \n"                            \
+    "vpackuswb   %%ymm0,%%ymm0,%%ymm0           \n"                            \
+    "vpackuswb   %%ymm1,%%ymm1,%%ymm1           \n"                            \
+    "vpackuswb   %%ymm2,%%ymm2,%%ymm2           \n"
+
+#if defined(HAS_I422TOBGRAROW_AVX2)
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 BGRA (64 bytes).
+void OMITFP I422ToBGRARow_AVX2(const uint8* y_buf,
+                               const uint8* u_buf,
+                               const uint8* v_buf,
+                               uint8* dst_bgra,
+                               int width) {
   asm volatile (
     "sub       %[u_buf],%[v_buf]               \n"
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "pxor      %%xmm4,%%xmm4                   \n"
+    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
     LABELALIGN
   "1:                                          \n"
-    READYUV422
-    YUVTORGB
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "punpcklbw %%xmm0,%%xmm1                   \n"
-    "punpcklbw %%xmm2,%%xmm5                   \n"
-    "movdqa    %%xmm5,%%xmm0                   \n"
-    "punpcklwd %%xmm1,%%xmm5                   \n"
-    "punpckhwd %%xmm1,%%xmm0                   \n"
-    "movdqu    %%xmm5," MEMACCESS([dst_bgra]) "\n"
-    "movdqu    %%xmm0," MEMACCESS2(0x10,[dst_bgra]) "\n"
-    "lea       " MEMLEA(0x20,[dst_bgra]) ",%[dst_bgra] \n"
-    "sub       $0x8,%[width]                   \n"
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+
+    // Step 3: Weave into BGRA
+    "vpunpcklbw %%ymm0,%%ymm1,%%ymm1           \n"  // GB
+    "vpermq     $0xd8,%%ymm1,%%ymm1            \n"
+    "vpunpcklbw %%ymm2,%%ymm5,%%ymm2           \n"  // AR
+    "vpermq     $0xd8,%%ymm2,%%ymm2            \n"
+    "vpunpcklwd %%ymm1,%%ymm2,%%ymm0           \n"  // ARGB first 8 pixels
+    "vpunpckhwd %%ymm1,%%ymm2,%%ymm2           \n"  // ARGB next 8 pixels
+
+    "vmovdqu    %%ymm0," MEMACCESS([dst_bgra]) "\n"
+    "vmovdqu    %%ymm2," MEMACCESS2(0x20,[dst_bgra]) "\n"
+    "lea       " MEMLEA(0x40,[dst_bgra]) ",%[dst_bgra] \n"
+    "sub       $0x10,%[width]                  \n"
     "jg        1b                              \n"
+    "vzeroupper                                \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
     [dst_bgra]"+r"(dst_bgra),  // %[dst_bgra]
     [width]"+rm"(width)    // %[width]
-  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB)  // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
   );
 }
+#endif  // HAS_I422TOBGRAROW_AVX2
 
-void OMITFP I422ToABGRRow_Unaligned_SSSE3(const uint8* y_buf,
-                                          const uint8* u_buf,
-                                          const uint8* v_buf,
-                                          uint8* dst_abgr,
-                                          int width) {
+#if defined(HAS_I422TOARGBROW_AVX2)
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+void OMITFP I422ToARGBRow_AVX2(const uint8* y_buf,
+                               const uint8* u_buf,
+                               const uint8* v_buf,
+                               uint8* dst_argb,
+                               int width) {
   asm volatile (
     "sub       %[u_buf],%[v_buf]               \n"
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "pxor      %%xmm4,%%xmm4                   \n"
+    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
     LABELALIGN
   "1:                                          \n"
-    READYUV422
-    YUVTORGB
-    "punpcklbw %%xmm1,%%xmm2                   \n"
-    "punpcklbw %%xmm5,%%xmm0                   \n"
-    "movdqa    %%xmm2,%%xmm1                   \n"
-    "punpcklwd %%xmm0,%%xmm2                   \n"
-    "punpckhwd %%xmm0,%%xmm1                   \n"
-    "movdqu    %%xmm2," MEMACCESS([dst_abgr]) "\n"
-    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_abgr]) "\n"
-    "lea       " MEMLEA(0x20,[dst_abgr]) ",%[dst_abgr] \n"
-    "sub       $0x8,%[width]                   \n"
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+
+    // Step 3: Weave into ARGB
+    "vpunpcklbw %%ymm1,%%ymm0,%%ymm0           \n"  // BG
+    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
+    "vpunpcklbw %%ymm5,%%ymm2,%%ymm2           \n"  // RA
+    "vpermq     $0xd8,%%ymm2,%%ymm2            \n"
+    "vpunpcklwd %%ymm2,%%ymm0,%%ymm1           \n"  // BGRA first 8 pixels
+    "vpunpckhwd %%ymm2,%%ymm0,%%ymm0           \n"  // BGRA next 8 pixels
+
+    "vmovdqu    %%ymm1," MEMACCESS([dst_argb]) "\n"
+    "vmovdqu    %%ymm0," MEMACCESS2(0x20,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x40,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x10,%[width]                  \n"
     "jg        1b                              \n"
+    "vzeroupper                                \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
-    [dst_abgr]"+r"(dst_abgr),  // %[dst_abgr]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
-  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB)  // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
   );
 }
+#endif  // HAS_I422TOARGBROW_AVX2
 
-void OMITFP I422ToRGBARow_Unaligned_SSSE3(const uint8* y_buf,
-                                          const uint8* u_buf,
-                                          const uint8* v_buf,
-                                          uint8* dst_rgba,
-                                          int width) {
+#if defined(HAS_I422TOABGRROW_AVX2)
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ABGR (64 bytes).
+void OMITFP I422ToABGRRow_AVX2(const uint8* y_buf,
+                               const uint8* u_buf,
+                               const uint8* v_buf,
+                               uint8* dst_argb,
+                               int width) {
   asm volatile (
     "sub       %[u_buf],%[v_buf]               \n"
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "pxor      %%xmm4,%%xmm4                   \n"
+    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
     LABELALIGN
   "1:                                          \n"
-    READYUV422
-    YUVTORGB
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "punpcklbw %%xmm2,%%xmm1                   \n"
-    "punpcklbw %%xmm0,%%xmm5                   \n"
-    "movdqa    %%xmm5,%%xmm0                   \n"
-    "punpcklwd %%xmm1,%%xmm5                   \n"
-    "punpckhwd %%xmm1,%%xmm0                   \n"
-    "movdqu    %%xmm5," MEMACCESS([dst_rgba]) "\n"
-    "movdqu    %%xmm0," MEMACCESS2(0x10,[dst_rgba]) "\n"
-    "lea       " MEMLEA(0x20,[dst_rgba]) ",%[dst_rgba] \n"
-    "sub       $0x8,%[width]                   \n"
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+
+    // Step 3: Weave into ABGR
+    "vpunpcklbw %%ymm1,%%ymm2,%%ymm1           \n"  // RG
+    "vpermq     $0xd8,%%ymm1,%%ymm1            \n"
+    "vpunpcklbw %%ymm5,%%ymm0,%%ymm2           \n"  // BA
+    "vpermq     $0xd8,%%ymm2,%%ymm2            \n"
+    "vpunpcklwd %%ymm2,%%ymm1,%%ymm0           \n"  // RGBA first 8 pixels
+    "vpunpckhwd %%ymm2,%%ymm1,%%ymm1           \n"  // RGBA next 8 pixels
+    "vmovdqu    %%ymm0," MEMACCESS([dst_argb]) "\n"
+    "vmovdqu    %%ymm1," MEMACCESS2(0x20,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x40,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x10,%[width]                  \n"
     "jg        1b                              \n"
+    "vzeroupper                                \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
-    [dst_rgba]"+r"(dst_rgba),  // %[dst_rgba]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
-  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB)  // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
   );
 }
+#endif  // HAS_I422TOABGRROW_AVX2
 
-#endif  // HAS_I422TOARGBROW_SSSE3
+#if defined(HAS_I422TORGBAROW_AVX2)
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 RGBA (64 bytes).
+void OMITFP I422ToRGBARow_AVX2(const uint8* y_buf,
+                               const uint8* u_buf,
+                               const uint8* v_buf,
+                               uint8* dst_argb,
+                               int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+
+    // Step 3: Weave into RGBA
+    "vpunpcklbw %%ymm2,%%ymm1,%%ymm1           \n"
+    "vpermq     $0xd8,%%ymm1,%%ymm1            \n"
+    "vpunpcklbw %%ymm0,%%ymm5,%%ymm2           \n"
+    "vpermq     $0xd8,%%ymm2,%%ymm2            \n"
+    "vpunpcklwd %%ymm1,%%ymm2,%%ymm0           \n"
+    "vpunpckhwd %%ymm1,%%ymm2,%%ymm1           \n"
+    "vmovdqu    %%ymm0," MEMACCESS([dst_argb]) "\n"
+    "vmovdqu    %%ymm1," MEMACCESS2(0x20,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x40,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x10,%[width]                  \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB)  // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+#endif  // HAS_I422TORGBAROW_AVX2
 
 #ifdef HAS_YTOARGBROW_SSE2
-void YToARGBRow_SSE2(const uint8* y_buf,
-                     uint8* dst_argb,
-                     int width) {
+void YToARGBRow_SSE2(const uint8* y_buf, uint8* dst_argb, int width) {
   asm volatile (
-    "pxor      %%xmm5,%%xmm5                   \n"
-    "pcmpeqb   %%xmm4,%%xmm4                   \n"
-    "pslld     $0x18,%%xmm4                    \n"
-    "mov       $0x00100010,%%eax               \n"
-    "movd      %%eax,%%xmm3                    \n"
-    "pshufd    $0x0,%%xmm3,%%xmm3              \n"
-    "mov       $0x004a004a,%%eax               \n"
+    "mov       $0x4a354a35,%%eax               \n"  // 4a35 = 18997 = 1.164
     "movd      %%eax,%%xmm2                    \n"
     "pshufd    $0x0,%%xmm2,%%xmm2              \n"
+    "mov       $0x04880488,%%eax               \n"  // 0488 = 1160 = 1.164 * 16
+    "movd      %%eax,%%xmm3                    \n"
+    "pshufd    $0x0,%%xmm3,%%xmm3              \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "pslld     $0x18,%%xmm4                    \n"
     LABELALIGN
   "1:                                          \n"
     // Step 1: Scale Y contribution to 8 G values. G = (y - 16) * 1.164
     "movq      " MEMACCESS(0) ",%%xmm0         \n"
     "lea       " MEMLEA(0x8,0) ",%0            \n"
-    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
     "psubusw   %%xmm3,%%xmm0                   \n"
-    "pmullw    %%xmm2,%%xmm0                   \n"
     "psrlw     $6, %%xmm0                      \n"
     "packuswb  %%xmm0,%%xmm0                   \n"
 
@@ -2939,8 +2095,8 @@ void YToARGBRow_SSE2(const uint8* y_buf,
     "punpckhwd %%xmm1,%%xmm1                   \n"
     "por       %%xmm4,%%xmm0                   \n"
     "por       %%xmm4,%%xmm1                   \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
-    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
     "lea       " MEMLEA(0x20,1) ",%1           \n"
 
     "sub       $0x8,%2                         \n"
@@ -2950,13 +2106,58 @@ void YToARGBRow_SSE2(const uint8* y_buf,
     "+rm"(width)     // %2
   :
   : "memory", "cc", "eax"
-#if defined(__SSE2__)
     , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
-#endif
   );
 }
 #endif  // HAS_YTOARGBROW_SSE2
 
+#ifdef HAS_YTOARGBROW_AVX2
+// 16 pixels of Y converted to 16 pixels of ARGB (64 bytes).
+// note: vpunpcklbw mutates and vpackuswb unmutates.
+void YToARGBRow_AVX2(const uint8* y_buf, uint8* dst_argb, int width) {
+  asm volatile (
+    "mov        $0x4a354a35,%%eax              \n" // 0488 = 1160 = 1.164 * 16
+    "vmovd      %%eax,%%xmm2                   \n"
+    "vbroadcastss %%xmm2,%%ymm2                \n"
+    "mov        $0x4880488,%%eax               \n" // 4a35 = 18997 = 1.164
+    "vmovd      %%eax,%%xmm3                   \n"
+    "vbroadcastss %%xmm3,%%ymm3                \n"
+    "vpcmpeqb   %%ymm4,%%ymm4,%%ymm4           \n"
+    "vpslld     $0x18,%%ymm4,%%ymm4            \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    // Step 1: Scale Y contribution to 16 G values. G = (y - 16) * 1.164
+    "vmovdqu    " MEMACCESS(0) ",%%xmm0        \n"
+    "lea        " MEMLEA(0x10,0) ",%0          \n"
+    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
+    "vpunpcklbw %%ymm0,%%ymm0,%%ymm0           \n"
+    "vpmulhuw   %%ymm2,%%ymm0,%%ymm0           \n"
+    "vpsubusw   %%ymm3,%%ymm0,%%ymm0           \n"
+    "vpsrlw     $0x6,%%ymm0,%%ymm0             \n"
+    "vpackuswb  %%ymm0,%%ymm0,%%ymm0           \n"
+    "vpunpcklbw %%ymm0,%%ymm0,%%ymm1           \n"
+    "vpermq     $0xd8,%%ymm1,%%ymm1            \n"
+    "vpunpcklwd %%ymm1,%%ymm1,%%ymm0           \n"
+    "vpunpckhwd %%ymm1,%%ymm1,%%ymm1           \n"
+    "vpor       %%ymm4,%%ymm0,%%ymm0           \n"
+    "vpor       %%ymm4,%%ymm1,%%ymm1           \n"
+    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
+    "vmovdqu    %%ymm1," MEMACCESS2(0x20,1) "  \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub        $0x10,%2                       \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(y_buf),     // %0
+    "+r"(dst_argb),  // %1
+    "+rm"(width)     // %2
+  :
+  : "memory", "cc", "eax"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+  );
+}
+#endif  // HAS_YTOARGBROW_AVX2
+
 #ifdef HAS_MIRRORROW_SSSE3
 // Shuffle table for reversing the bytes.
 static uvec8 kShuffleMirror = {
@@ -2967,38 +2168,56 @@ void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
   intptr_t temp_width = (intptr_t)(width);
   asm volatile (
     "movdqa    %3,%%xmm5                       \n"
-    "lea       " MEMLEA(-0x10,0) ",%0          \n"
     LABELALIGN
   "1:                                          \n"
-    MEMOPREG(movdqa,0x00,0,2,1,xmm0)           //  movdqa  (%0,%2),%%xmm0
+    MEMOPREG(movdqu,-0x10,0,2,1,xmm0)          //  movdqu -0x10(%0,%2),%%xmm0
     "pshufb    %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
     "jg        1b                              \n"
   : "+r"(src),  // %0
     "+r"(dst),  // %1
     "+r"(temp_width)  // %2
   : "m"(kShuffleMirror) // %3
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm5"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm5"
   );
 }
 #endif  // HAS_MIRRORROW_SSSE3
 
+#ifdef HAS_MIRRORROW_AVX2
+void MirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    "vbroadcastf128 %3,%%ymm5                  \n"
+    LABELALIGN
+  "1:                                          \n"
+    MEMOPREG(vmovdqu,-0x20,0,2,1,ymm0)         //  vmovdqu -0x20(%0,%2),%%ymm0
+    "vpshufb    %%ymm5,%%ymm0,%%ymm0           \n"
+    "vpermq     $0x4e,%%ymm0,%%ymm0            \n"
+    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(temp_width)  // %2
+  : "m"(kShuffleMirror) // %3
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm5"
+  );
+}
+#endif  // HAS_MIRRORROW_AVX2
+
 #ifdef HAS_MIRRORROW_SSE2
 void MirrorRow_SSE2(const uint8* src, uint8* dst, int width) {
   intptr_t temp_width = (intptr_t)(width);
   asm volatile (
-    "lea       " MEMLEA(-0x10,0) ",%0          \n"
     LABELALIGN
   "1:                                          \n"
-    MEMOPREG(movdqu,0x00,0,2,1,xmm0)           //  movdqu  (%0,%2),%%xmm0
+    MEMOPREG(movdqu,-0x10,0,2,1,xmm0)          //  movdqu -0x10(%0,%2),%%xmm0
     "movdqa    %%xmm0,%%xmm1                   \n"
     "psllw     $0x8,%%xmm0                     \n"
     "psrlw     $0x8,%%xmm1                     \n"
@@ -3006,21 +2225,16 @@ void MirrorRow_SSE2(const uint8* src, uint8* dst, int width) {
     "pshuflw   $0x1b,%%xmm0,%%xmm0             \n"
     "pshufhw   $0x1b,%%xmm0,%%xmm0             \n"
     "pshufd    $0x4e,%%xmm0,%%xmm0             \n"
-    "sub       $0x10,%2                        \n"
     "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x10,1)",%1            \n"
+    "sub       $0x10,%2                        \n"
     "jg        1b                              \n"
   : "+r"(src),  // %0
     "+r"(dst),  // %1
     "+r"(temp_width)  // %2
   :
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1"
   );
 }
 #endif  // HAS_MIRRORROW_SSE2
@@ -3035,108 +2249,119 @@ void MirrorUVRow_SSSE3(const uint8* src, uint8* dst_u, uint8* dst_v,
   intptr_t temp_width = (intptr_t)(width);
   asm volatile (
     "movdqa    %4,%%xmm1                       \n"
-    "lea       " MEMLEA4(-0x10,0,3,2) ",%0       \n"
+    "lea       " MEMLEA4(-0x10,0,3,2) ",%0     \n"
     "sub       %1,%2                           \n"
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "lea       " MEMLEA(-0x10,0) ",%0            \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(-0x10,0) ",%0          \n"
     "pshufb    %%xmm1,%%xmm0                   \n"
-    "sub       $8,%3                           \n"
     "movlpd    %%xmm0," MEMACCESS(1) "         \n"
-    BUNDLEALIGN
     MEMOPMEM(movhpd,xmm0,0x00,1,2,1)           //  movhpd    %%xmm0,(%1,%2)
     "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $8,%3                           \n"
     "jg        1b                              \n"
   : "+r"(src),      // %0
     "+r"(dst_u),    // %1
     "+r"(dst_v),    // %2
     "+r"(temp_width)  // %3
   : "m"(kShuffleMirrorUV)  // %4
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1"
   );
 }
 #endif  // HAS_MIRRORROW_UV_SSSE3
 
-#ifdef HAS_ARGBMIRRORROW_SSSE3
-// Shuffle table for reversing the bytes.
-static uvec8 kARGBShuffleMirror = {
-  12u, 13u, 14u, 15u, 8u, 9u, 10u, 11u, 4u, 5u, 6u, 7u, 0u, 1u, 2u, 3u
-};
+#ifdef HAS_ARGBMIRRORROW_SSE2
 
-void ARGBMirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
+void ARGBMirrorRow_SSE2(const uint8* src, uint8* dst, int width) {
   intptr_t temp_width = (intptr_t)(width);
   asm volatile (
     "lea       " MEMLEA4(-0x10,0,2,4) ",%0     \n"
-    "movdqa    %3,%%xmm5                       \n"
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "pshufb    %%xmm5,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "pshufd    $0x1b,%%xmm0,%%xmm0             \n"
     "lea       " MEMLEA(-0x10,0) ",%0          \n"
-    "sub       $0x4,%2                         \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%2                         \n"
     "jg        1b                              \n"
   : "+r"(src),  // %0
     "+r"(dst),  // %1
     "+r"(temp_width)  // %2
-  : "m"(kARGBShuffleMirror)  // %3
+  :
   : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm5"
-#endif
+    , "xmm0"
   );
 }
-#endif  // HAS_ARGBMIRRORROW_SSSE3
+#endif  // HAS_ARGBMIRRORROW_SSE2
 
-#ifdef HAS_SPLITUVROW_SSE2
-void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) {
+#ifdef HAS_ARGBMIRRORROW_AVX2
+// Shuffle table for reversing the bytes.
+static const ulvec32 kARGBShuffleMirror_AVX2 = {
+  7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
+};
+void ARGBMirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
+  intptr_t temp_width = (intptr_t)(width);
   asm volatile (
-    "pcmpeqb    %%xmm5,%%xmm5                    \n"
-    "psrlw      $0x8,%%xmm5                      \n"
+    "vmovdqu    %3,%%ymm5                      \n"
+    LABELALIGN
+  "1:                                          \n"
+    VMEMOPREG(vpermd,-0x20,0,2,4,ymm5,ymm0) // vpermd -0x20(%0,%2,4),ymm5,ymm0
+    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
+    "lea        " MEMLEA(0x20,1) ",%1          \n"
+    "sub        $0x8,%2                        \n"
+    "jg         1b                             \n"
+    "vzeroupper                                \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(temp_width)  // %2
+  : "m"(kARGBShuffleMirror_AVX2) // %3
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm5"
+  );
+}
+#endif  // HAS_ARGBMIRRORROW_AVX2
+
+#ifdef HAS_SPLITUVROW_AVX2
+void SplitUVRow_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5             \n"
+    "vpsrlw     $0x8,%%ymm5,%%ymm5               \n"
     "sub        %1,%2                            \n"
     LABELALIGN
   "1:                                            \n"
-    "movdqa     " MEMACCESS(0) ",%%xmm0          \n"
-    "movdqa     " MEMACCESS2(0x10,0) ",%%xmm1    \n"
-    "lea        " MEMLEA(0x20,0) ",%0            \n"
-    "movdqa     %%xmm0,%%xmm2                    \n"
-    "movdqa     %%xmm1,%%xmm3                    \n"
-    "pand       %%xmm5,%%xmm0                    \n"
-    "pand       %%xmm5,%%xmm1                    \n"
-    "packuswb   %%xmm1,%%xmm0                    \n"
-    "psrlw      $0x8,%%xmm2                      \n"
-    "psrlw      $0x8,%%xmm3                      \n"
-    "packuswb   %%xmm3,%%xmm2                    \n"
-    "movdqa     %%xmm0," MEMACCESS(1) "          \n"
-    MEMOPMEM(movdqa,xmm2,0x00,1,2,1)             // movdqa     %%xmm2,(%1,%2)
-    "lea        " MEMLEA(0x10,1) ",%1            \n"
-    "sub        $0x10,%3                         \n"
+    "vmovdqu    " MEMACCESS(0) ",%%ymm0          \n"
+    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1    \n"
+    "lea        " MEMLEA(0x40,0) ",%0            \n"
+    "vpsrlw     $0x8,%%ymm0,%%ymm2               \n"
+    "vpsrlw     $0x8,%%ymm1,%%ymm3               \n"
+    "vpand      %%ymm5,%%ymm0,%%ymm0             \n"
+    "vpand      %%ymm5,%%ymm1,%%ymm1             \n"
+    "vpackuswb  %%ymm1,%%ymm0,%%ymm0             \n"
+    "vpackuswb  %%ymm3,%%ymm2,%%ymm2             \n"
+    "vpermq     $0xd8,%%ymm0,%%ymm0              \n"
+    "vpermq     $0xd8,%%ymm2,%%ymm2              \n"
+    "vmovdqu    %%ymm0," MEMACCESS(1) "          \n"
+    MEMOPMEM(vmovdqu,ymm2,0x00,1,2,1)             //  vmovdqu %%ymm2,(%1,%2)
+    "lea        " MEMLEA(0x20,1) ",%1            \n"
+    "sub        $0x20,%3                         \n"
     "jg         1b                               \n"
+    "vzeroupper                                  \n"
   : "+r"(src_uv),     // %0
     "+r"(dst_u),      // %1
     "+r"(dst_v),      // %2
     "+r"(pix)         // %3
   :
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
   );
 }
+#endif  // HAS_SPLITUVROW_AVX2
 
-void SplitUVRow_Unaligned_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
-                               int pix) {
+#ifdef HAS_SPLITUVROW_SSE2
+void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) {
   asm volatile (
     "pcmpeqb    %%xmm5,%%xmm5                    \n"
     "psrlw      $0x8,%%xmm5                      \n"
@@ -3164,52 +2389,46 @@ void SplitUVRow_Unaligned_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
     "+r"(dst_v),      // %2
     "+r"(pix)         // %3
   :
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
   );
 }
 #endif  // HAS_SPLITUVROW_SSE2
 
-#ifdef HAS_MERGEUVROW_SSE2
-void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+#ifdef HAS_MERGEUVROW_AVX2
+void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
                      int width) {
   asm volatile (
     "sub       %0,%1                             \n"
     LABELALIGN
   "1:                                            \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0           \n"
-    MEMOPREG(movdqa,0x00,0,1,1,xmm1)             //  movdqa    (%0,%1,1),%%xmm1
-    "lea       " MEMLEA(0x10,0) ",%0             \n"
-    "movdqa    %%xmm0,%%xmm2                     \n"
-    "punpcklbw %%xmm1,%%xmm0                     \n"
-    "punpckhbw %%xmm1,%%xmm2                     \n"
-    "movdqa    %%xmm0," MEMACCESS(2) "           \n"
-    "movdqa    %%xmm2," MEMACCESS2(0x10,2) "     \n"
-    "lea       " MEMLEA(0x20,2) ",%2             \n"
-    "sub       $0x10,%3                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0           \n"
+    MEMOPREG(vmovdqu,0x00,0,1,1,ymm1)             //  vmovdqu (%0,%1,1),%%ymm1
+    "lea       " MEMLEA(0x20,0) ",%0             \n"
+    "vpunpcklbw %%ymm1,%%ymm0,%%ymm2             \n"
+    "vpunpckhbw %%ymm1,%%ymm0,%%ymm0             \n"
+    "vextractf128 $0x0,%%ymm2," MEMACCESS(2) "   \n"
+    "vextractf128 $0x0,%%ymm0," MEMACCESS2(0x10,2) "\n"
+    "vextractf128 $0x1,%%ymm2," MEMACCESS2(0x20,2) "\n"
+    "vextractf128 $0x1,%%ymm0," MEMACCESS2(0x30,2) "\n"
+    "lea       " MEMLEA(0x40,2) ",%2             \n"
+    "sub       $0x20,%3                          \n"
     "jg        1b                                \n"
+    "vzeroupper                                  \n"
   : "+r"(src_u),     // %0
     "+r"(src_v),     // %1
     "+r"(dst_uv),    // %2
     "+r"(width)      // %3
   :
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2"
   );
 }
+#endif  // HAS_MERGEUVROW_AVX2
 
-void MergeUVRow_Unaligned_SSE2(const uint8* src_u, const uint8* src_v,
-                               uint8* dst_uv, int width) {
+#ifdef HAS_MERGEUVROW_SSE2
+void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
   asm volatile (
     "sub       %0,%1                             \n"
     LABELALIGN
@@ -3230,13 +2449,8 @@ void MergeUVRow_Unaligned_SSE2(const uint8* src_u, const uint8* src_v,
     "+r"(dst_uv),    // %2
     "+r"(width)      // %3
   :
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2"
   );
 }
 #endif  // HAS_MERGEUVROW_SSE2
@@ -3246,11 +2460,11 @@ void CopyRow_SSE2(const uint8* src, uint8* dst, int count) {
   asm volatile (
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
     "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
-    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
     "lea       " MEMLEA(0x20,1) ",%1           \n"
     "sub       $0x20,%2                        \n"
     "jg        1b                              \n"
@@ -3259,30 +2473,36 @@ void CopyRow_SSE2(const uint8* src, uint8* dst, int count) {
     "+r"(count)  // %2
   :
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1"
-#endif
   );
 }
 #endif  // HAS_COPYROW_SSE2
 
-#ifdef HAS_COPYROW_X86
-void CopyRow_X86(const uint8* src, uint8* dst, int width) {
-  size_t width_tmp = (size_t)(width);
+#ifdef HAS_COPYROW_AVX
+void CopyRow_AVX(const uint8* src, uint8* dst, int count) {
   asm volatile (
-    "shr       $0x2,%2                         \n"
-    "rep movsl " MEMMOVESTRING(0,1) "          \n"
-  : "+S"(src),  // %0
-    "+D"(dst),  // %1
-    "+c"(width_tmp) // %2
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vmovdqu   %%ymm0," MEMACCESS(1) "         \n"
+    "vmovdqu   %%ymm1," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x40,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(count)  // %2
   :
   : "memory", "cc"
+    , "xmm0", "xmm1"
   );
 }
-#endif  // HAS_COPYROW_X86
+#endif  // HAS_COPYROW_AVX
 
 #ifdef HAS_COPYROW_ERMS
-// Unaligned Multiple of 1.
+// Multiple of 1.
 void CopyRow_ERMS(const uint8* src, uint8* dst, int width) {
   size_t width_tmp = (size_t)(width);
   asm volatile (
@@ -3306,19 +2526,19 @@ void ARGBCopyAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
     "psrld     $0x8,%%xmm1                     \n"
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm2         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm3   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm3   \n"
     "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "movdqa    " MEMACCESS(1) ",%%xmm4         \n"
-    "movdqa    " MEMACCESS2(0x10,1) ",%%xmm5   \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm4         \n"
+    "movdqu    " MEMACCESS2(0x10,1) ",%%xmm5   \n"
     "pand      %%xmm0,%%xmm2                   \n"
     "pand      %%xmm0,%%xmm3                   \n"
     "pand      %%xmm1,%%xmm4                   \n"
     "pand      %%xmm1,%%xmm5                   \n"
     "por       %%xmm4,%%xmm2                   \n"
     "por       %%xmm5,%%xmm3                   \n"
-    "movdqa    %%xmm2," MEMACCESS(1) "         \n"
-    "movdqa    %%xmm3," MEMACCESS2(0x10,1) "   \n"
+    "movdqu    %%xmm2," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm3," MEMACCESS2(0x10,1) "   \n"
     "lea       " MEMLEA(0x20,1) ",%1           \n"
     "sub       $0x8,%2                         \n"
     "jg        1b                              \n"
@@ -3327,9 +2547,7 @@ void ARGBCopyAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
     "+r"(width)  // %2
   :
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
   );
 }
 #endif  // HAS_ARGBCOPYALPHAROW_SSE2
@@ -3358,9 +2576,7 @@ void ARGBCopyAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
     "+r"(width)  // %2
   :
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1", "xmm2"
-#endif
   );
 }
 #endif  // HAS_ARGBCOPYALPHAROW_AVX2
@@ -3380,16 +2596,16 @@ void ARGBCopyYToAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
     "punpcklbw %%xmm2,%%xmm2                   \n"
     "punpckhwd %%xmm2,%%xmm3                   \n"
     "punpcklwd %%xmm2,%%xmm2                   \n"
-    "movdqa    " MEMACCESS(1) ",%%xmm4         \n"
-    "movdqa    " MEMACCESS2(0x10,1) ",%%xmm5   \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm4         \n"
+    "movdqu    " MEMACCESS2(0x10,1) ",%%xmm5   \n"
     "pand      %%xmm0,%%xmm2                   \n"
     "pand      %%xmm0,%%xmm3                   \n"
     "pand      %%xmm1,%%xmm4                   \n"
     "pand      %%xmm1,%%xmm5                   \n"
     "por       %%xmm4,%%xmm2                   \n"
     "por       %%xmm5,%%xmm3                   \n"
-    "movdqa    %%xmm2," MEMACCESS(1) "         \n"
-    "movdqa    %%xmm3," MEMACCESS2(0x10,1) "   \n"
+    "movdqu    %%xmm2," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm3," MEMACCESS2(0x10,1) "   \n"
     "lea       " MEMLEA(0x20,1) ",%1           \n"
     "sub       $0x8,%2                         \n"
     "jg        1b                              \n"
@@ -3398,9 +2614,7 @@ void ARGBCopyYToAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
     "+r"(width)  // %2
   :
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
   );
 }
 #endif  // HAS_ARGBCOPYYTOALPHAROW_SSE2
@@ -3431,18 +2645,16 @@ void ARGBCopyYToAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
     "+r"(width)  // %2
   :
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1", "xmm2"
-#endif
   );
 }
 #endif  // HAS_ARGBCOPYYTOALPHAROW_AVX2
 
 #ifdef HAS_SETROW_X86
-void SetRow_X86(uint8* dst, uint32 v32, int width) {
-  size_t width_tmp = (size_t)(width);
+void SetRow_X86(uint8* dst, uint8 v8, int width) {
+  size_t width_tmp = (size_t)(width >> 2);
+  const uint32 v32 = v8 * 0x01010101;  // Duplicate byte to all bytes.
   asm volatile (
-    "shr       $0x2,%1                         \n"
     "rep stosl " MEMSTORESTRING(eax,0) "       \n"
     : "+D"(dst),       // %0
       "+c"(width_tmp)  // %1
@@ -3450,19 +2662,24 @@ void SetRow_X86(uint8* dst, uint32 v32, int width) {
     : "memory", "cc");
 }
 
-void ARGBSetRows_X86(uint8* dst, uint32 v32, int width,
-                   int dst_stride, int height) {
-  for (int y = 0; y < height; ++y) {
-    size_t width_tmp = (size_t)(width);
-    uint32* d = (uint32*)(dst);
-    asm volatile (
-      "rep stosl " MEMSTORESTRING(eax,0) "     \n"
-      : "+D"(d),         // %0
-        "+c"(width_tmp)  // %1
-      : "a"(v32)         // %2
-      : "memory", "cc");
-    dst += dst_stride;
-  }
+void SetRow_ERMS(uint8* dst, uint8 v8, int width) {
+  size_t width_tmp = (size_t)(width);
+  asm volatile (
+    "rep stosb " MEMSTORESTRING(al,0) "        \n"
+    : "+D"(dst),       // %0
+      "+c"(width_tmp)  // %1
+    : "a"(v8)          // %2
+    : "memory", "cc");
+}
+
+void ARGBSetRow_X86(uint8* dst_argb, uint32 v32, int width) {
+  size_t width_tmp = (size_t)(width);
+  asm volatile (
+    "rep stosl " MEMSTORESTRING(eax,0) "       \n"
+    : "+D"(dst_argb),  // %0
+      "+c"(width_tmp)  // %1
+    : "a"(v32)         // %2
+    : "memory", "cc");
 }
 #endif  // HAS_SETROW_X86
 
@@ -3473,13 +2690,13 @@ void YUY2ToYRow_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix) {
     "psrlw     $0x8,%%xmm5                     \n"
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
     "lea       " MEMLEA(0x20,0) ",%0           \n"
     "pand      %%xmm5,%%xmm0                   \n"
     "pand      %%xmm5,%%xmm1                   \n"
     "packuswb  %%xmm1,%%xmm0                   \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x10,1) ",%1           \n"
     "sub       $0x10,%2                        \n"
     "jg        1b                              \n"
@@ -3488,9 +2705,7 @@ void YUY2ToYRow_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix) {
     "+r"(pix)        // %2
   :
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1", "xmm5"
-#endif
   );
 }
 
@@ -3502,11 +2717,10 @@ void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
     "sub       %1,%2                           \n"
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    BUNDLEALIGN
-    MEMOPREG(movdqa,0x00,0,4,1,xmm2)           //  movdqa  (%0,%4,1),%%xmm2
-    MEMOPREG(movdqa,0x10,0,4,1,xmm3)           //  movdqa  0x10(%0,%4,1),%%xmm3
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x00,0,4,1,xmm2)           //  movdqu  (%0,%4,1),%%xmm2
+    MEMOPREG(movdqu,0x10,0,4,1,xmm3)           //  movdqu  0x10(%0,%4,1),%%xmm3
     "lea       " MEMLEA(0x20,0) ",%0           \n"
     "pavgb     %%xmm2,%%xmm0                   \n"
     "pavgb     %%xmm3,%%xmm1                   \n"
@@ -3519,7 +2733,6 @@ void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
     "psrlw     $0x8,%%xmm1                     \n"
     "packuswb  %%xmm1,%%xmm1                   \n"
     "movq      %%xmm0," MEMACCESS(1) "         \n"
-    BUNDLEALIGN
     MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
     "lea       " MEMLEA(0x8,1) ",%1            \n"
     "sub       $0x10,%3                        \n"
@@ -3529,13 +2742,8 @@ void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
     "+r"(dst_v),       // %2
     "+r"(pix)          // %3
   : "r"((intptr_t)(stride_yuy2))  // %4
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
   );
 }
 
@@ -3547,8 +2755,8 @@ void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
     "sub       %1,%2                           \n"
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
     "lea       " MEMLEA(0x20,0) ",%0           \n"
     "psrlw     $0x8,%%xmm0                     \n"
     "psrlw     $0x8,%%xmm1                     \n"
@@ -3559,7 +2767,6 @@ void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
     "psrlw     $0x8,%%xmm1                     \n"
     "packuswb  %%xmm1,%%xmm1                   \n"
     "movq      %%xmm0," MEMACCESS(1) "         \n"
-    BUNDLEALIGN
     MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
     "lea       " MEMLEA(0x8,1) ",%1            \n"
     "sub       $0x10,%3                        \n"
@@ -3569,47 +2776,36 @@ void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
     "+r"(dst_v),       // %2
     "+r"(pix)          // %3
   :
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm5"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm5"
   );
 }
 
-void YUY2ToYRow_Unaligned_SSE2(const uint8* src_yuy2,
-                               uint8* dst_y, int pix) {
+void UYVYToYRow_SSE2(const uint8* src_uyvy, uint8* dst_y, int pix) {
   asm volatile (
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "psrlw     $0x8,%%xmm5                     \n"
     LABELALIGN
   "1:                                          \n"
     "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
     "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
     "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "pand      %%xmm5,%%xmm0                   \n"
-    "pand      %%xmm5,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
     "packuswb  %%xmm1,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
     "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
     "jg        1b                              \n"
-  : "+r"(src_yuy2),  // %0
+  : "+r"(src_uyvy),  // %0
     "+r"(dst_y),     // %1
     "+r"(pix)        // %2
   :
   : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm5"
-#endif
+    , "xmm0", "xmm1"
   );
 }
 
-void YUY2ToUVRow_Unaligned_SSE2(const uint8* src_yuy2,
-                                int stride_yuy2,
-                                uint8* dst_u, uint8* dst_v, int pix) {
+void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
   asm volatile (
     "pcmpeqb   %%xmm5,%%xmm5                   \n"
     "psrlw     $0x8,%%xmm5                     \n"
@@ -3618,14 +2814,13 @@ void YUY2ToUVRow_Unaligned_SSE2(const uint8* src_yuy2,
   "1:                                          \n"
     "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
     "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    BUNDLEALIGN
     MEMOPREG(movdqu,0x00,0,4,1,xmm2)           //  movdqu  (%0,%4,1),%%xmm2
     MEMOPREG(movdqu,0x10,0,4,1,xmm3)           //  movdqu  0x10(%0,%4,1),%%xmm3
     "lea       " MEMLEA(0x20,0) ",%0           \n"
     "pavgb     %%xmm2,%%xmm0                   \n"
     "pavgb     %%xmm3,%%xmm1                   \n"
-    "psrlw     $0x8,%%xmm0                     \n"
-    "psrlw     $0x8,%%xmm1                     \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
     "packuswb  %%xmm1,%%xmm0                   \n"
     "movdqa    %%xmm0,%%xmm1                   \n"
     "pand      %%xmm5,%%xmm0                   \n"
@@ -3633,28 +2828,22 @@ void YUY2ToUVRow_Unaligned_SSE2(const uint8* src_yuy2,
     "psrlw     $0x8,%%xmm1                     \n"
     "packuswb  %%xmm1,%%xmm1                   \n"
     "movq      %%xmm0," MEMACCESS(1) "         \n"
-    BUNDLEALIGN
     MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
     "lea       " MEMLEA(0x8,1) ",%1            \n"
     "sub       $0x10,%3                        \n"
     "jg        1b                              \n"
-  : "+r"(src_yuy2),    // %0
+  : "+r"(src_uyvy),    // %0
     "+r"(dst_u),       // %1
     "+r"(dst_v),       // %2
     "+r"(pix)          // %3
-  : "r"((intptr_t)(stride_yuy2))  // %4
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-#endif
+  : "r"((intptr_t)(stride_uyvy))  // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
   );
 }
 
-void YUY2ToUV422Row_Unaligned_SSE2(const uint8* src_yuy2,
-                                   uint8* dst_u, uint8* dst_v, int pix) {
+void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) {
   asm volatile (
     "pcmpeqb   %%xmm5,%%xmm5                   \n"
     "psrlw     $0x8,%%xmm5                     \n"
@@ -3664,8 +2853,8 @@ void YUY2ToUV422Row_Unaligned_SSE2(const uint8* src_yuy2,
     "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
     "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
     "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "psrlw     $0x8,%%xmm0                     \n"
-    "psrlw     $0x8,%%xmm1                     \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
     "packuswb  %%xmm1,%%xmm0                   \n"
     "movdqa    %%xmm0,%%xmm1                   \n"
     "pand      %%xmm5,%%xmm0                   \n"
@@ -3673,247 +2862,226 @@ void YUY2ToUV422Row_Unaligned_SSE2(const uint8* src_yuy2,
     "psrlw     $0x8,%%xmm1                     \n"
     "packuswb  %%xmm1,%%xmm1                   \n"
     "movq      %%xmm0," MEMACCESS(1) "         \n"
-    BUNDLEALIGN
     MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
     "lea       " MEMLEA(0x8,1) ",%1            \n"
     "sub       $0x10,%3                        \n"
     "jg        1b                              \n"
-  : "+r"(src_yuy2),    // %0
+  : "+r"(src_uyvy),    // %0
     "+r"(dst_u),       // %1
     "+r"(dst_v),       // %2
     "+r"(pix)          // %3
   :
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm5"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm5"
   );
 }
+#endif  // HAS_YUY2TOYROW_SSE2
 
-void UYVYToYRow_SSE2(const uint8* src_uyvy, uint8* dst_y, int pix) {
+#ifdef HAS_YUY2TOYROW_AVX2
+void YUY2ToYRow_AVX2(const uint8* src_yuy2, uint8* dst_y, int pix) {
   asm volatile (
+    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
+    "vpsrlw    $0x8,%%ymm5,%%ymm5              \n"
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "psrlw     $0x8,%%xmm0                     \n"
-    "psrlw     $0x8,%%xmm1                     \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpand     %%ymm5,%%ymm0,%%ymm0            \n"
+    "vpand     %%ymm5,%%ymm1,%%ymm1            \n"
+    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vmovdqu   %%ymm0," MEMACCESS(1) "         \n"
+    "lea      " MEMLEA(0x20,1) ",%1            \n"
+    "sub       $0x20,%2                        \n"
     "jg        1b                              \n"
-  : "+r"(src_uyvy),  // %0
+    "vzeroupper                                \n"
+  : "+r"(src_yuy2),  // %0
     "+r"(dst_y),     // %1
     "+r"(pix)        // %2
   :
   : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1"
-#endif
+    , "xmm0", "xmm1", "xmm5"
   );
 }
 
-void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
+void YUY2ToUVRow_AVX2(const uint8* src_yuy2, int stride_yuy2,
                       uint8* dst_u, uint8* dst_v, int pix) {
   asm volatile (
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "psrlw     $0x8,%%xmm5                     \n"
+    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
+    "vpsrlw    $0x8,%%ymm5,%%ymm5              \n"
     "sub       %1,%2                           \n"
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    BUNDLEALIGN
-    MEMOPREG(movdqa,0x00,0,4,1,xmm2)           //  movdqa  (%0,%4,1),%%xmm2
-    MEMOPREG(movdqa,0x10,0,4,1,xmm3)           //  movdqa  0x10(%0,%4,1),%%xmm3
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "pavgb     %%xmm2,%%xmm0                   \n"
-    "pavgb     %%xmm3,%%xmm1                   \n"
-    "pand      %%xmm5,%%xmm0                   \n"
-    "pand      %%xmm5,%%xmm1                   \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "pand      %%xmm5,%%xmm0                   \n"
-    "packuswb  %%xmm0,%%xmm0                   \n"
-    "psrlw     $0x8,%%xmm1                     \n"
-    "packuswb  %%xmm1,%%xmm1                   \n"
-    "movq      %%xmm0," MEMACCESS(1) "         \n"
-    BUNDLEALIGN
-    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "sub       $0x10,%3                        \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    VMEMOPREG(vpavgb,0x00,0,4,1,ymm0,ymm0)     // vpavgb (%0,%4,1),%%ymm0,%%ymm0
+    VMEMOPREG(vpavgb,0x20,0,4,1,ymm1,ymm1)
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
+    "vpsrlw    $0x8,%%ymm1,%%ymm1              \n"
+    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vpand     %%ymm5,%%ymm0,%%ymm1            \n"
+    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
+    "vpackuswb %%ymm1,%%ymm1,%%ymm1            \n"
+    "vpackuswb %%ymm0,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm1,%%ymm1             \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vextractf128 $0x0,%%ymm1," MEMACCESS(1) " \n"
+    VEXTOPMEM(vextractf128,0,ymm0,0x00,1,2,1) // vextractf128 $0x0,%%ymm0,(%1,%2,1)
+    "lea      " MEMLEA(0x10,1) ",%1            \n"
+    "sub       $0x20,%3                        \n"
     "jg        1b                              \n"
-  : "+r"(src_uyvy),    // %0
+    "vzeroupper                                \n"
+  : "+r"(src_yuy2),    // %0
     "+r"(dst_u),       // %1
     "+r"(dst_v),       // %2
     "+r"(pix)          // %3
-  : "r"((intptr_t)(stride_uyvy))  // %4
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-#endif
+  : "r"((intptr_t)(stride_yuy2))  // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm5"
   );
 }
 
-void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
+void YUY2ToUV422Row_AVX2(const uint8* src_yuy2,
                          uint8* dst_u, uint8* dst_v, int pix) {
   asm volatile (
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "psrlw     $0x8,%%xmm5                     \n"
+    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
+    "vpsrlw    $0x8,%%ymm5,%%ymm5              \n"
     "sub       %1,%2                           \n"
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "pand      %%xmm5,%%xmm0                   \n"
-    "pand      %%xmm5,%%xmm1                   \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "pand      %%xmm5,%%xmm0                   \n"
-    "packuswb  %%xmm0,%%xmm0                   \n"
-    "psrlw     $0x8,%%xmm1                     \n"
-    "packuswb  %%xmm1,%%xmm1                   \n"
-    "movq      %%xmm0," MEMACCESS(1) "         \n"
-    BUNDLEALIGN
-    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "sub       $0x10,%3                        \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
+    "vpsrlw    $0x8,%%ymm1,%%ymm1              \n"
+    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vpand     %%ymm5,%%ymm0,%%ymm1            \n"
+    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
+    "vpackuswb %%ymm1,%%ymm1,%%ymm1            \n"
+    "vpackuswb %%ymm0,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm1,%%ymm1             \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vextractf128 $0x0,%%ymm1," MEMACCESS(1) " \n"
+    VEXTOPMEM(vextractf128,0,ymm0,0x00,1,2,1) // vextractf128 $0x0,%%ymm0,(%1,%2,1)
+    "lea      " MEMLEA(0x10,1) ",%1            \n"
+    "sub       $0x20,%3                        \n"
     "jg        1b                              \n"
-  : "+r"(src_uyvy),    // %0
+    "vzeroupper                                \n"
+  : "+r"(src_yuy2),    // %0
     "+r"(dst_u),       // %1
     "+r"(dst_v),       // %2
     "+r"(pix)          // %3
   :
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm5"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm5"
   );
 }
 
-void UYVYToYRow_Unaligned_SSE2(const uint8* src_uyvy,
-                               uint8* dst_y, int pix) {
+void UYVYToYRow_AVX2(const uint8* src_uyvy, uint8* dst_y, int pix) {
   asm volatile (
     LABELALIGN
   "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "psrlw     $0x8,%%xmm0                     \n"
-    "psrlw     $0x8,%%xmm1                     \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
+    "vpsrlw    $0x8,%%ymm1,%%ymm1              \n"
+    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vmovdqu   %%ymm0," MEMACCESS(1) "         \n"
+    "lea      " MEMLEA(0x20,1) ",%1            \n"
+    "sub       $0x20,%2                        \n"
     "jg        1b                              \n"
+    "vzeroupper                                \n"
   : "+r"(src_uyvy),  // %0
     "+r"(dst_y),     // %1
     "+r"(pix)        // %2
   :
   : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1"
-#endif
+    , "xmm0", "xmm1", "xmm5"
   );
 }
-
-void UYVYToUVRow_Unaligned_SSE2(const uint8* src_uyvy, int stride_uyvy,
-                                uint8* dst_u, uint8* dst_v, int pix) {
+void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
   asm volatile (
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "psrlw     $0x8,%%xmm5                     \n"
+    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
+    "vpsrlw    $0x8,%%ymm5,%%ymm5              \n"
     "sub       %1,%2                           \n"
+
     LABELALIGN
   "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    BUNDLEALIGN
-    MEMOPREG(movdqu,0x00,0,4,1,xmm2)           //  movdqu  (%0,%4,1),%%xmm2
-    MEMOPREG(movdqu,0x10,0,4,1,xmm3)           //  movdqu  0x10(%0,%4,1),%%xmm3
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "pavgb     %%xmm2,%%xmm0                   \n"
-    "pavgb     %%xmm3,%%xmm1                   \n"
-    "pand      %%xmm5,%%xmm0                   \n"
-    "pand      %%xmm5,%%xmm1                   \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "pand      %%xmm5,%%xmm0                   \n"
-    "packuswb  %%xmm0,%%xmm0                   \n"
-    "psrlw     $0x8,%%xmm1                     \n"
-    "packuswb  %%xmm1,%%xmm1                   \n"
-    "movq      %%xmm0," MEMACCESS(1) "         \n"
-    BUNDLEALIGN
-    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "sub       $0x10,%3                        \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    VMEMOPREG(vpavgb,0x00,0,4,1,ymm0,ymm0)     // vpavgb (%0,%4,1),%%ymm0,%%ymm0
+    VMEMOPREG(vpavgb,0x20,0,4,1,ymm1,ymm1)
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpand     %%ymm5,%%ymm0,%%ymm0            \n"
+    "vpand     %%ymm5,%%ymm1,%%ymm1            \n"
+    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vpand     %%ymm5,%%ymm0,%%ymm1            \n"
+    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
+    "vpackuswb %%ymm1,%%ymm1,%%ymm1            \n"
+    "vpackuswb %%ymm0,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm1,%%ymm1             \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vextractf128 $0x0,%%ymm1," MEMACCESS(1) " \n"
+    VEXTOPMEM(vextractf128,0,ymm0,0x00,1,2,1) // vextractf128 $0x0,%%ymm0,(%1,%2,1)
+    "lea      " MEMLEA(0x10,1) ",%1            \n"
+    "sub       $0x20,%3                        \n"
     "jg        1b                              \n"
+    "vzeroupper                                \n"
   : "+r"(src_uyvy),    // %0
     "+r"(dst_u),       // %1
     "+r"(dst_v),       // %2
     "+r"(pix)          // %3
   : "r"((intptr_t)(stride_uyvy))  // %4
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm5"
   );
 }
 
-void UYVYToUV422Row_Unaligned_SSE2(const uint8* src_uyvy,
-                                   uint8* dst_u, uint8* dst_v, int pix) {
+void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) {
   asm volatile (
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "psrlw     $0x8,%%xmm5                     \n"
+    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+    "vpsrlw     $0x8,%%ymm5,%%ymm5             \n"
     "sub       %1,%2                           \n"
     LABELALIGN
   "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "pand      %%xmm5,%%xmm0                   \n"
-    "pand      %%xmm5,%%xmm1                   \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "pand      %%xmm5,%%xmm0                   \n"
-    "packuswb  %%xmm0,%%xmm0                   \n"
-    "psrlw     $0x8,%%xmm1                     \n"
-    "packuswb  %%xmm1,%%xmm1                   \n"
-    "movq      %%xmm0," MEMACCESS(1) "         \n"
-    BUNDLEALIGN
-    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "sub       $0x10,%3                        \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpand     %%ymm5,%%ymm0,%%ymm0            \n"
+    "vpand     %%ymm5,%%ymm1,%%ymm1            \n"
+    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vpand     %%ymm5,%%ymm0,%%ymm1            \n"
+    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
+    "vpackuswb %%ymm1,%%ymm1,%%ymm1            \n"
+    "vpackuswb %%ymm0,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm1,%%ymm1             \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vextractf128 $0x0,%%ymm1," MEMACCESS(1) " \n"
+    VEXTOPMEM(vextractf128,0,ymm0,0x00,1,2,1) // vextractf128 $0x0,%%ymm0,(%1,%2,1)
+    "lea      " MEMLEA(0x10,1) ",%1            \n"
+    "sub       $0x20,%3                        \n"
     "jg        1b                              \n"
+    "vzeroupper                                \n"
   : "+r"(src_uyvy),    // %0
     "+r"(dst_u),       // %1
     "+r"(dst_v),       // %2
     "+r"(pix)          // %3
   :
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm5"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm5"
   );
 }
-#endif  // HAS_YUY2TOYROW_SSE2
+#endif  // HAS_YUY2TOYROW_AVX2
 
 #ifdef HAS_ARGBBLENDROW_SSE2
 // Blend 8 pixels at a time.
@@ -3956,9 +3124,9 @@ void ARGBBlendRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
     "paddusb   %%xmm2,%%xmm0                   \n"
     "pand      %%xmm5,%%xmm1                   \n"
     "paddusb   %%xmm1,%%xmm0                   \n"
-    "sub       $0x1,%3                         \n"
     "movd      %%xmm0," MEMACCESS(2) "         \n"
     "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "sub       $0x1,%3                         \n"
     "jge       10b                             \n"
 
   "19:                                         \n"
@@ -3988,9 +3156,9 @@ void ARGBBlendRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
     "paddusb   %%xmm2,%%xmm0                   \n"
     "pand      %%xmm5,%%xmm1                   \n"
     "paddusb   %%xmm1,%%xmm0                   \n"
-    "sub       $0x4,%3                         \n"
-    "movdqa    %%xmm0," MEMACCESS(2) "         \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
     "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
     "jge       41b                             \n"
 
   "49:                                         \n"
@@ -4019,9 +3187,9 @@ void ARGBBlendRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
     "paddusb   %%xmm2,%%xmm0                   \n"
     "pand      %%xmm5,%%xmm1                   \n"
     "paddusb   %%xmm1,%%xmm0                   \n"
-    "sub       $0x1,%3                         \n"
     "movd      %%xmm0," MEMACCESS(2) "         \n"
     "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "sub       $0x1,%3                         \n"
     "jge       91b                             \n"
   "99:                                         \n"
   : "+r"(src_argb0),    // %0
@@ -4030,9 +3198,7 @@ void ARGBBlendRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
     "+r"(width)         // %3
   :
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-#endif
   );
 }
 #endif  // HAS_ARGBBLENDROW_SSE2
@@ -4091,49 +3257,18 @@ void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
     "paddusb   %%xmm2,%%xmm0                   \n"
     "pand      %%xmm5,%%xmm1                   \n"
     "paddusb   %%xmm1,%%xmm0                   \n"
-    "sub       $0x1,%3                         \n"
     "movd      %%xmm0," MEMACCESS(2) "         \n"
     "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "sub       $0x1,%3                         \n"
     "jge       10b                             \n"
 
   "19:                                         \n"
     "add       $1-4,%3                         \n"
     "jl        49f                             \n"
-    "test      $0xf,%0                         \n"
-    "jne       41f                             \n"
-    "test      $0xf,%1                         \n"
-    "jne       41f                             \n"
 
     // 4 pixel loop.
     LABELALIGN
   "40:                                         \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm3         \n"
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "movdqa    %%xmm3,%%xmm0                   \n"
-    "pxor      %%xmm4,%%xmm3                   \n"
-    "movdqa    " MEMACCESS(1) ",%%xmm2         \n"
-    "pshufb    %4,%%xmm3                       \n"
-    "pand      %%xmm6,%%xmm2                   \n"
-    "paddw     %%xmm7,%%xmm3                   \n"
-    "pmullw    %%xmm3,%%xmm2                   \n"
-    "movdqa    " MEMACCESS(1) ",%%xmm1         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "psrlw     $0x8,%%xmm1                     \n"
-    "por       %%xmm4,%%xmm0                   \n"
-    "pmullw    %%xmm3,%%xmm1                   \n"
-    "psrlw     $0x8,%%xmm2                     \n"
-    "paddusb   %%xmm2,%%xmm0                   \n"
-    "pand      %%xmm5,%%xmm1                   \n"
-    "paddusb   %%xmm1,%%xmm0                   \n"
-    "sub       $0x4,%3                         \n"
-    "movdqa    %%xmm0," MEMACCESS(2) "         \n"
-    "lea       " MEMLEA(0x10,2) ",%2           \n"
-    "jge       40b                             \n"
-    "jmp       49f                             \n"
-
-    // 4 pixel unaligned loop.
-    LABELALIGN
-  "41:                                         \n"
     "movdqu    " MEMACCESS(0) ",%%xmm3         \n"
     "lea       " MEMLEA(0x10,0) ",%0           \n"
     "movdqa    %%xmm3,%%xmm0                   \n"
@@ -4152,10 +3287,10 @@ void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
     "paddusb   %%xmm2,%%xmm0                   \n"
     "pand      %%xmm5,%%xmm1                   \n"
     "paddusb   %%xmm1,%%xmm0                   \n"
-    "sub       $0x4,%3                         \n"
-    "movdqa    %%xmm0," MEMACCESS(2) "         \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
     "lea       " MEMLEA(0x10,2) ",%2           \n"
-    "jge       41b                             \n"
+    "sub       $0x4,%3                         \n"
+    "jge       40b                             \n"
 
   "49:                                         \n"
     "add       $0x3,%3                         \n"
@@ -4181,9 +3316,9 @@ void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
     "paddusb   %%xmm2,%%xmm0                   \n"
     "pand      %%xmm5,%%xmm1                   \n"
     "paddusb   %%xmm1,%%xmm0                   \n"
-    "sub       $0x1,%3                         \n"
     "movd      %%xmm0," MEMACCESS(2) "         \n"
     "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "sub       $0x1,%3                         \n"
     "jge       91b                             \n"
   "99:                                         \n"
   : "+r"(src_argb0),    // %0
@@ -4192,16 +3327,13 @@ void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
     "+r"(width)         // %3
   : "m"(kShuffleAlpha)  // %4
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-#endif
   );
 }
 #endif  // HAS_ARGBBLENDROW_SSSE3
 
 #ifdef HAS_ARGBATTENUATEROW_SSE2
 // Attenuate 4 pixels at a time.
-// aligned to 16 bytes
 void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width) {
   asm volatile (
     "pcmpeqb   %%xmm4,%%xmm4                   \n"
@@ -4212,17 +3344,17 @@ void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width) {
     // 4 pixel loop.
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
     "punpcklbw %%xmm0,%%xmm0                   \n"
     "pshufhw   $0xff,%%xmm0,%%xmm2             \n"
     "pshuflw   $0xff,%%xmm2,%%xmm2             \n"
     "pmulhuw   %%xmm2,%%xmm0                   \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm1         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
     "punpckhbw %%xmm1,%%xmm1                   \n"
     "pshufhw   $0xff,%%xmm1,%%xmm2             \n"
     "pshuflw   $0xff,%%xmm2,%%xmm2             \n"
     "pmulhuw   %%xmm2,%%xmm1                   \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm2         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
     "lea       " MEMLEA(0x10,0) ",%0           \n"
     "psrlw     $0x8,%%xmm0                     \n"
     "pand      %%xmm4,%%xmm2                   \n"
@@ -4230,18 +3362,16 @@ void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width) {
     "packuswb  %%xmm1,%%xmm0                   \n"
     "pand      %%xmm5,%%xmm0                   \n"
     "por       %%xmm2,%%xmm0                   \n"
-    "sub       $0x4,%2                         \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%2                         \n"
     "jg        1b                              \n"
   : "+r"(src_argb),    // %0
     "+r"(dst_argb),    // %1
     "+r"(width)        // %2
   :
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
   );
 }
 #endif  // HAS_ARGBATTENUATEROW_SSE2
@@ -4249,14 +3379,13 @@ void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width) {
 #ifdef HAS_ARGBATTENUATEROW_SSSE3
 // Shuffle table duplicating alpha
 static uvec8 kShuffleAlpha0 = {
-  3u, 3u, 3u, 3u, 3u, 3u, 128u, 128u, 7u, 7u, 7u, 7u, 7u, 7u, 128u, 128u,
+  3u, 3u, 3u, 3u, 3u, 3u, 128u, 128u, 7u, 7u, 7u, 7u, 7u, 7u, 128u, 128u
 };
 static uvec8 kShuffleAlpha1 = {
   11u, 11u, 11u, 11u, 11u, 11u, 128u, 128u,
-  15u, 15u, 15u, 15u, 15u, 15u, 128u, 128u,
+  15u, 15u, 15u, 15u, 15u, 15u, 128u, 128u
 };
 // Attenuate 4 pixels at a time.
-// aligned to 16 bytes
 void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
   asm volatile (
     "pcmpeqb   %%xmm3,%%xmm3                   \n"
@@ -4284,9 +3413,9 @@ void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
     "psrlw     $0x8,%%xmm1                     \n"
     "packuswb  %%xmm1,%%xmm0                   \n"
     "por       %%xmm2,%%xmm0                   \n"
-    "sub       $0x4,%2                         \n"
     "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%2                         \n"
     "jg        1b                              \n"
   : "+r"(src_argb),    // %0
     "+r"(dst_argb),    // %1
@@ -4294,16 +3423,56 @@ void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
   : "m"(kShuffleAlpha0),  // %3
     "m"(kShuffleAlpha1)  // %4
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
   );
 }
 #endif  // HAS_ARGBATTENUATEROW_SSSE3
 
+#ifdef HAS_ARGBATTENUATEROW_AVX2
+// Shuffle table duplicating alpha.
+static const uvec8 kShuffleAlpha_AVX2 = {
+  6u, 7u, 6u, 7u, 6u, 7u, 128u, 128u, 14u, 15u, 14u, 15u, 14u, 15u, 128u, 128u
+};
+// Attenuate 8 pixels at a time.
+void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "vbroadcastf128 %3,%%ymm4                  \n"
+    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+    "vpslld     $0x18,%%ymm5,%%ymm5            \n"
+    "sub        %0,%1                          \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu    " MEMACCESS(0) ",%%ymm6        \n"
+    "vpunpcklbw %%ymm6,%%ymm6,%%ymm0           \n"
+    "vpunpckhbw %%ymm6,%%ymm6,%%ymm1           \n"
+    "vpshufb    %%ymm4,%%ymm0,%%ymm2           \n"
+    "vpshufb    %%ymm4,%%ymm1,%%ymm3           \n"
+    "vpmulhuw   %%ymm2,%%ymm0,%%ymm0           \n"
+    "vpmulhuw   %%ymm3,%%ymm1,%%ymm1           \n"
+    "vpand      %%ymm5,%%ymm6,%%ymm6           \n"
+    "vpsrlw     $0x8,%%ymm0,%%ymm0             \n"
+    "vpsrlw     $0x8,%%ymm1,%%ymm1             \n"
+    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
+    "vpor       %%ymm6,%%ymm0,%%ymm0           \n"
+    MEMOPMEM(vmovdqu,ymm0,0x00,0,1,1)          //  vmovdqu %%ymm0,(%0,%1)
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "sub        $0x8,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(width)        // %2
+  : "m"(kShuffleAlpha_AVX2)  // %3
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+  );
+}
+#endif  // HAS_ARGBATTENUATEROW_AVX2
+
 #ifdef HAS_ARGBUNATTENUATEROW_SSE2
 // Unattenuate 4 pixels at a time.
-// aligned to 16 bytes
 void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb,
                              int width) {
   uintptr_t alpha = 0;
@@ -4324,7 +3493,6 @@ void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb,
     "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
     "movzb     " MEMACCESS2(0x0b,0) ",%3       \n"
     "punpckhbw %%xmm1,%%xmm1                   \n"
-    BUNDLEALIGN
     MEMOPREG(movd,0x00,4,3,4,xmm2)             //  movd      0x0(%4,%3,4),%%xmm2
     "movzb     " MEMACCESS2(0x0f,0) ",%3       \n"
     MEMOPREG(movd,0x00,4,3,4,xmm3)             //  movd      0x0(%4,%3,4),%%xmm3
@@ -4334,26 +3502,90 @@ void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb,
     "pmulhuw   %%xmm2,%%xmm1                   \n"
     "lea       " MEMLEA(0x10,0) ",%0           \n"
     "packuswb  %%xmm1,%%xmm0                   \n"
-    "sub       $0x4,%2                         \n"
     "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%2                         \n"
     "jg        1b                              \n"
   : "+r"(src_argb),    // %0
     "+r"(dst_argb),    // %1
     "+r"(width),       // %2
     "+r"(alpha)        // %3
   : "r"(fixed_invtbl8)  // %4
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
 }
 #endif  // HAS_ARGBUNATTENUATEROW_SSE2
 
+#ifdef HAS_ARGBUNATTENUATEROW_AVX2
+// Shuffle table duplicating alpha.
+static const uvec8 kUnattenShuffleAlpha_AVX2 = {
+  0u, 1u, 0u, 1u, 0u, 1u, 6u, 7u, 8u, 9u, 8u, 9u, 8u, 9u, 14u, 15u
+};
+// Unattenuate 8 pixels at a time.
+void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                             int width) {
+  uintptr_t alpha = 0;
+  asm volatile (
+    "sub        %0,%1                          \n"
+    "vbroadcastf128 %5,%%ymm5                  \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    // replace VPGATHER
+    "movzb     " MEMACCESS2(0x03,0) ",%3       \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm0)             //  vmovd 0x0(%4,%3,4),%%xmm0
+    "movzb     " MEMACCESS2(0x07,0) ",%3       \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm1)             //  vmovd 0x0(%4,%3,4),%%xmm1
+    "movzb     " MEMACCESS2(0x0b,0) ",%3       \n"
+    "vpunpckldq %%xmm1,%%xmm0,%%xmm6           \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm2)             //  vmovd 0x0(%4,%3,4),%%xmm2
+    "movzb     " MEMACCESS2(0x0f,0) ",%3       \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm3)             //  vmovd 0x0(%4,%3,4),%%xmm3
+    "movzb     " MEMACCESS2(0x13,0) ",%3       \n"
+    "vpunpckldq %%xmm3,%%xmm2,%%xmm7           \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm0)             //  vmovd 0x0(%4,%3,4),%%xmm0
+    "movzb     " MEMACCESS2(0x17,0) ",%3       \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm1)             //  vmovd 0x0(%4,%3,4),%%xmm1
+    "movzb     " MEMACCESS2(0x1b,0) ",%3       \n"
+    "vpunpckldq %%xmm1,%%xmm0,%%xmm0           \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm2)             //  vmovd 0x0(%4,%3,4),%%xmm2
+    "movzb     " MEMACCESS2(0x1f,0) ",%3       \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm3)             //  vmovd 0x0(%4,%3,4),%%xmm3
+    "vpunpckldq %%xmm3,%%xmm2,%%xmm2           \n"
+    "vpunpcklqdq %%xmm7,%%xmm6,%%xmm3          \n"
+    "vpunpcklqdq %%xmm2,%%xmm0,%%xmm0          \n"
+    "vinserti128 $0x1,%%xmm0,%%ymm3,%%ymm3     \n"
+    // end of VPGATHER
+
+    "vmovdqu    " MEMACCESS(0) ",%%ymm6        \n"
+    "vpunpcklbw %%ymm6,%%ymm6,%%ymm0           \n"
+    "vpunpckhbw %%ymm6,%%ymm6,%%ymm1           \n"
+    "vpunpcklwd %%ymm3,%%ymm3,%%ymm2           \n"
+    "vpunpckhwd %%ymm3,%%ymm3,%%ymm3           \n"
+    "vpshufb    %%ymm5,%%ymm2,%%ymm2           \n"
+    "vpshufb    %%ymm5,%%ymm3,%%ymm3           \n"
+    "vpmulhuw   %%ymm2,%%ymm0,%%ymm0           \n"
+    "vpmulhuw   %%ymm3,%%ymm1,%%ymm1           \n"
+    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
+    MEMOPMEM(vmovdqu,ymm0,0x00,0,1,1)          //  vmovdqu %%ymm0,(%0,%1)
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "sub        $0x8,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(width),       // %2
+    "+r"(alpha)        // %3
+  : "r"(fixed_invtbl8),  // %4
+    "m"(kUnattenShuffleAlpha_AVX2)  // %5
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_ARGBUNATTENUATEROW_AVX2
+
 #ifdef HAS_ARGBGRAYROW_SSSE3
 // Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels
 void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
@@ -4364,16 +3596,16 @@ void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
     // 8 pixel loop.
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
     "pmaddubsw %%xmm4,%%xmm0                   \n"
     "pmaddubsw %%xmm4,%%xmm1                   \n"
     "phaddw    %%xmm1,%%xmm0                   \n"
     "paddw     %%xmm5,%%xmm0                   \n"
     "psrlw     $0x7,%%xmm0                     \n"
     "packuswb  %%xmm0,%%xmm0                   \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm2         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm3   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm3   \n"
     "lea       " MEMLEA(0x20,0) ",%0           \n"
     "psrld     $0x18,%%xmm2                    \n"
     "psrld     $0x18,%%xmm3                    \n"
@@ -4385,10 +3617,10 @@ void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
     "movdqa    %%xmm0,%%xmm1                   \n"
     "punpcklwd %%xmm3,%%xmm0                   \n"
     "punpckhwd %%xmm3,%%xmm1                   \n"
-    "sub       $0x8,%2                         \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
-    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
     "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
     "jg        1b                              \n"
   : "+r"(src_argb),   // %0
     "+r"(dst_argb),   // %1
@@ -4396,9 +3628,7 @@ void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
   : "m"(kARGBToYJ),   // %3
     "m"(kAddYJ64)     // %4
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
   );
 }
 #endif  // HAS_ARGBGRAYROW_SSSE3
@@ -4430,30 +3660,30 @@ void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width) {
     // 8 pixel loop.
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
     "pmaddubsw %%xmm2,%%xmm0                   \n"
     "pmaddubsw %%xmm2,%%xmm6                   \n"
     "phaddw    %%xmm6,%%xmm0                   \n"
     "psrlw     $0x7,%%xmm0                     \n"
     "packuswb  %%xmm0,%%xmm0                   \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm5         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm5         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
     "pmaddubsw %%xmm3,%%xmm5                   \n"
     "pmaddubsw %%xmm3,%%xmm1                   \n"
     "phaddw    %%xmm1,%%xmm5                   \n"
     "psrlw     $0x7,%%xmm5                     \n"
     "packuswb  %%xmm5,%%xmm5                   \n"
     "punpcklbw %%xmm5,%%xmm0                   \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm5         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm5         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
     "pmaddubsw %%xmm4,%%xmm5                   \n"
     "pmaddubsw %%xmm4,%%xmm1                   \n"
     "phaddw    %%xmm1,%%xmm5                   \n"
     "psrlw     $0x7,%%xmm5                     \n"
     "packuswb  %%xmm5,%%xmm5                   \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm6         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm6         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
     "psrld     $0x18,%%xmm6                    \n"
     "psrld     $0x18,%%xmm1                    \n"
     "packuswb  %%xmm1,%%xmm6                   \n"
@@ -4462,10 +3692,10 @@ void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width) {
     "movdqa    %%xmm0,%%xmm1                   \n"
     "punpcklwd %%xmm5,%%xmm0                   \n"
     "punpckhwd %%xmm5,%%xmm1                   \n"
-    "sub       $0x8,%1                         \n"
-    "movdqa    %%xmm0," MEMACCESS(0) "         \n"
-    "movdqa    %%xmm1," MEMACCESS2(0x10,0) "   \n"
+    "movdqu    %%xmm0," MEMACCESS(0) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,0) "   \n"
     "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "sub       $0x8,%1                         \n"
     "jg        1b                              \n"
   : "+r"(dst_argb),      // %0
     "+r"(width)          // %1
@@ -4473,9 +3703,7 @@ void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width) {
     "m"(kARGBToSepiaG),  // %3
     "m"(kARGBToSepiaR)   // %4
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
-#endif
   );
 }
 #endif  // HAS_ARGBSEPIAROW_SSSE3
@@ -4495,12 +3723,12 @@ void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
     // 8 pixel loop.
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm7   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm7   \n"
     "pmaddubsw %%xmm2,%%xmm0                   \n"
     "pmaddubsw %%xmm2,%%xmm7                   \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm6         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm6         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
     "pmaddubsw %%xmm3,%%xmm6                   \n"
     "pmaddubsw %%xmm3,%%xmm1                   \n"
     "phaddsw   %%xmm7,%%xmm0                   \n"
@@ -4510,13 +3738,13 @@ void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
     "packuswb  %%xmm0,%%xmm0                   \n"
     "packuswb  %%xmm6,%%xmm6                   \n"
     "punpcklbw %%xmm6,%%xmm0                   \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm1         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm7   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm7   \n"
     "pmaddubsw %%xmm4,%%xmm1                   \n"
     "pmaddubsw %%xmm4,%%xmm7                   \n"
     "phaddsw   %%xmm7,%%xmm1                   \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm6         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm7   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm6         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm7   \n"
     "pmaddubsw %%xmm5,%%xmm6                   \n"
     "pmaddubsw %%xmm5,%%xmm7                   \n"
     "phaddsw   %%xmm7,%%xmm6                   \n"
@@ -4528,27 +3756,24 @@ void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
     "movdqa    %%xmm0,%%xmm6                   \n"
     "punpcklwd %%xmm1,%%xmm0                   \n"
     "punpckhwd %%xmm1,%%xmm6                   \n"
-    "sub       $0x8,%2                         \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
-    "movdqa    %%xmm6," MEMACCESS2(0x10,1) "   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm6," MEMACCESS2(0x10,1) "   \n"
     "lea       " MEMLEA(0x20,0) ",%0           \n"
     "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
     "jg        1b                              \n"
   : "+r"(src_argb),      // %0
     "+r"(dst_argb),      // %1
     "+r"(width)          // %2
   : "r"(matrix_argb)     // %3
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-#endif
   );
 }
 #endif  // HAS_ARGBCOLORMATRIXROW_SSSE3
 
 #ifdef HAS_ARGBQUANTIZEROW_SSE2
 // Quantize 4 ARGB pixels (16 bytes).
-// aligned to 16 bytes
 void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
                           int interval_offset, int width) {
   asm volatile (
@@ -4568,23 +3793,23 @@ void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
     // 4 pixel loop.
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
     "punpcklbw %%xmm5,%%xmm0                   \n"
     "pmulhuw   %%xmm2,%%xmm0                   \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm1         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
     "punpckhbw %%xmm5,%%xmm1                   \n"
     "pmulhuw   %%xmm2,%%xmm1                   \n"
     "pmullw    %%xmm3,%%xmm0                   \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm7         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm7         \n"
     "pmullw    %%xmm3,%%xmm1                   \n"
     "pand      %%xmm6,%%xmm7                   \n"
     "paddw     %%xmm4,%%xmm0                   \n"
     "paddw     %%xmm4,%%xmm1                   \n"
     "packuswb  %%xmm1,%%xmm0                   \n"
     "por       %%xmm7,%%xmm0                   \n"
-    "sub       $0x4,%1                         \n"
-    "movdqa    %%xmm0," MEMACCESS(0) "         \n"
+    "movdqu    %%xmm0," MEMACCESS(0) "         \n"
     "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "sub       $0x4,%1                         \n"
     "jg        1b                              \n"
   : "+r"(dst_argb),       // %0
     "+r"(width)           // %1
@@ -4592,16 +3817,13 @@ void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
     "r"(interval_size),   // %3
     "r"(interval_offset)  // %4
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-#endif
   );
 }
 #endif  // HAS_ARGBQUANTIZEROW_SSE2
 
 #ifdef HAS_ARGBSHADEROW_SSE2
 // Shade 4 pixels at a time by specified value.
-// Aligned to 16 bytes.
 void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
                        uint32 value) {
   asm volatile (
@@ -4612,7 +3834,7 @@ void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
     // 4 pixel loop.
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
     "lea       " MEMLEA(0x10,0) ",%0           \n"
     "movdqa    %%xmm0,%%xmm1                   \n"
     "punpcklbw %%xmm0,%%xmm0                   \n"
@@ -4622,18 +3844,16 @@ void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
     "psrlw     $0x8,%%xmm0                     \n"
     "psrlw     $0x8,%%xmm1                     \n"
     "packuswb  %%xmm1,%%xmm0                   \n"
-    "sub       $0x4,%2                         \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%2                         \n"
     "jg        1b                              \n"
   : "+r"(src_argb),  // %0
     "+r"(dst_argb),  // %1
     "+r"(width)      // %2
   : "r"(value)       // %3
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1", "xmm2"
-#endif
   );
 }
 #endif  // HAS_ARGBSHADEROW_SSE2
@@ -4643,7 +3863,7 @@ void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
 void ARGBMultiplyRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
                           uint8* dst_argb, int width) {
   asm volatile (
-    "pxor      %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm5,%%xmm5                  \n"
 
     // 4 pixel loop.
     LABELALIGN
@@ -4661,9 +3881,9 @@ void ARGBMultiplyRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
     "pmulhuw   %%xmm2,%%xmm0                   \n"
     "pmulhuw   %%xmm3,%%xmm1                   \n"
     "packuswb  %%xmm1,%%xmm0                   \n"
-    "sub       $0x4,%3                         \n"
     "movdqu    %%xmm0," MEMACCESS(2) "         \n"
     "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
     "jg        1b                              \n"
   : "+r"(src_argb0),  // %0
     "+r"(src_argb1),  // %1
@@ -4671,13 +3891,50 @@ void ARGBMultiplyRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
     "+r"(width)       // %3
   :
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-#endif
   );
 }
 #endif  // HAS_ARGBMULTIPLYROW_SSE2
 
+#ifdef HAS_ARGBMULTIPLYROW_AVX2
+// Multiply 2 rows of ARGB pixels together, 8 pixels at a time.
+void ARGBMultiplyRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    "vpxor      %%ymm5,%%ymm5,%%ymm5           \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu    " MEMACCESS(0) ",%%ymm1        \n"
+    "lea        " MEMLEA(0x20,0) ",%0          \n"
+    "vmovdqu    " MEMACCESS(1) ",%%ymm3        \n"
+    "lea        " MEMLEA(0x20,1) ",%1          \n"
+    "vpunpcklbw %%ymm1,%%ymm1,%%ymm0           \n"
+    "vpunpckhbw %%ymm1,%%ymm1,%%ymm1           \n"
+    "vpunpcklbw %%ymm5,%%ymm3,%%ymm2           \n"
+    "vpunpckhbw %%ymm5,%%ymm3,%%ymm3           \n"
+    "vpmulhuw   %%ymm2,%%ymm0,%%ymm0           \n"
+    "vpmulhuw   %%ymm3,%%ymm1,%%ymm1           \n"
+    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
+    "vmovdqu    %%ymm0," MEMACCESS(2) "        \n"
+    "lea       " MEMLEA(0x20,2) ",%2           \n"
+    "sub        $0x8,%3                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+#if defined(__AVX2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBMULTIPLYROW_AVX2
+
 #ifdef HAS_ARGBADDROW_SSE2
 // Add 2 rows of ARGB pixels together, 4 pixels at a time.
 void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
@@ -4691,9 +3948,9 @@ void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
     "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
     "lea       " MEMLEA(0x10,1) ",%1           \n"
     "paddusb   %%xmm1,%%xmm0                   \n"
-    "sub       $0x4,%3                         \n"
     "movdqu    %%xmm0," MEMACCESS(2) "         \n"
     "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
     "jg        1b                              \n"
   : "+r"(src_argb0),  // %0
     "+r"(src_argb1),  // %1
@@ -4701,13 +3958,39 @@ void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
     "+r"(width)       // %3
   :
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1"
-#endif
   );
 }
 #endif  // HAS_ARGBADDROW_SSE2
 
+#ifdef HAS_ARGBADDROW_AVX2
+// Add 2 rows of ARGB pixels together, 4 pixels at a time.
+void ARGBAddRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
+    "lea        " MEMLEA(0x20,0) ",%0          \n"
+    "vpaddusb   " MEMACCESS(1) ",%%ymm0,%%ymm0 \n"
+    "lea        " MEMLEA(0x20,1) ",%1          \n"
+    "vmovdqu    %%ymm0," MEMACCESS(2) "        \n"
+    "lea        " MEMLEA(0x20,2) ",%2          \n"
+    "sub        $0x8,%3                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+    , "xmm0"
+  );
+}
+#endif  // HAS_ARGBADDROW_AVX2
+
 #ifdef HAS_ARGBSUBTRACTROW_SSE2
 // Subtract 2 rows of ARGB pixels, 4 pixels at a time.
 void ARGBSubtractRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
@@ -4721,9 +4004,9 @@ void ARGBSubtractRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
     "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
     "lea       " MEMLEA(0x10,1) ",%1           \n"
     "psubusb   %%xmm1,%%xmm0                   \n"
-    "sub       $0x4,%3                         \n"
     "movdqu    %%xmm0," MEMACCESS(2) "         \n"
     "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
     "jg        1b                              \n"
   : "+r"(src_argb0),  // %0
     "+r"(src_argb1),  // %1
@@ -4731,13 +4014,39 @@ void ARGBSubtractRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
     "+r"(width)       // %3
   :
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1"
-#endif
   );
 }
 #endif  // HAS_ARGBSUBTRACTROW_SSE2
 
+#ifdef HAS_ARGBSUBTRACTROW_AVX2
+// Subtract 2 rows of ARGB pixels, 8 pixels at a time.
+void ARGBSubtractRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
+    "lea        " MEMLEA(0x20,0) ",%0          \n"
+    "vpsubusb   " MEMACCESS(1) ",%%ymm0,%%ymm0 \n"
+    "lea        " MEMLEA(0x20,1) ",%1          \n"
+    "vmovdqu    %%ymm0," MEMACCESS(2) "        \n"
+    "lea        " MEMLEA(0x20,2) ",%2          \n"
+    "sub        $0x8,%3                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+    , "xmm0"
+  );
+}
+#endif  // HAS_ARGBSUBTRACTROW_AVX2
+
 #ifdef HAS_SOBELXROW_SSE2
 // SobelX as a matrix is
 // -1  0  1
@@ -4759,13 +4068,11 @@ void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
     "punpcklbw %%xmm5,%%xmm0                   \n"
     "punpcklbw %%xmm5,%%xmm1                   \n"
     "psubw     %%xmm1,%%xmm0                   \n"
-    BUNDLEALIGN
     MEMOPREG(movq,0x00,0,1,1,xmm1)             //  movq      (%0,%1,1),%%xmm1
     MEMOPREG(movq,0x02,0,1,1,xmm2)             //  movq      0x2(%0,%1,1),%%xmm2
     "punpcklbw %%xmm5,%%xmm1                   \n"
     "punpcklbw %%xmm5,%%xmm2                   \n"
     "psubw     %%xmm2,%%xmm1                   \n"
-    BUNDLEALIGN
     MEMOPREG(movq,0x00,0,2,1,xmm2)             //  movq      (%0,%2,1),%%xmm2
     MEMOPREG(movq,0x02,0,2,1,xmm3)             //  movq      0x2(%0,%2,1),%%xmm3
     "punpcklbw %%xmm5,%%xmm2                   \n"
@@ -4778,10 +4085,9 @@ void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
     "psubw     %%xmm0,%%xmm1                   \n"
     "pmaxsw    %%xmm1,%%xmm0                   \n"
     "packuswb  %%xmm0,%%xmm0                   \n"
-    "sub       $0x8,%4                         \n"
-    BUNDLEALIGN
     MEMOPMEM(movq,xmm0,0x00,0,3,1)             //  movq      %%xmm0,(%0,%3,1)
     "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "sub       $0x8,%4                         \n"
     "jg        1b                              \n"
   : "+r"(src_y0),      // %0
     "+r"(src_y1),      // %1
@@ -4789,13 +4095,8 @@ void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
     "+r"(dst_sobelx),  // %3
     "+r"(width)        // %4
   :
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
   );
 }
 #endif  // HAS_SOBELXROW_SSE2
@@ -4820,13 +4121,11 @@ void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
     "punpcklbw %%xmm5,%%xmm0                   \n"
     "punpcklbw %%xmm5,%%xmm1                   \n"
     "psubw     %%xmm1,%%xmm0                   \n"
-    BUNDLEALIGN
     "movq      " MEMACCESS2(0x1,0) ",%%xmm1    \n"
     MEMOPREG(movq,0x01,0,1,1,xmm2)             //  movq      0x1(%0,%1,1),%%xmm2
     "punpcklbw %%xmm5,%%xmm1                   \n"
     "punpcklbw %%xmm5,%%xmm2                   \n"
     "psubw     %%xmm2,%%xmm1                   \n"
-    BUNDLEALIGN
     "movq      " MEMACCESS2(0x2,0) ",%%xmm2    \n"
     MEMOPREG(movq,0x02,0,1,1,xmm3)             //  movq      0x2(%0,%1,1),%%xmm3
     "punpcklbw %%xmm5,%%xmm2                   \n"
@@ -4839,23 +4138,17 @@ void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
     "psubw     %%xmm0,%%xmm1                   \n"
     "pmaxsw    %%xmm1,%%xmm0                   \n"
     "packuswb  %%xmm0,%%xmm0                   \n"
-    "sub       $0x8,%3                         \n"
-    BUNDLEALIGN
     MEMOPMEM(movq,xmm0,0x00,0,2,1)             //  movq      %%xmm0,(%0,%2,1)
     "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "sub       $0x8,%3                         \n"
     "jg        1b                              \n"
   : "+r"(src_y0),      // %0
     "+r"(src_y1),      // %1
     "+r"(dst_sobely),  // %2
     "+r"(width)        // %3
   :
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
   );
 }
 #endif  // HAS_SOBELYROW_SSE2
@@ -4876,8 +4169,8 @@ void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
     // 8 pixel loop.
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    MEMOPREG(movdqa,0x00,0,1,1,xmm1)           //  movdqa    (%0,%1,1),%%xmm1
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,1,1,xmm1)           //  movdqu    (%0,%1,1),%%xmm1
     "lea       " MEMLEA(0x10,0) ",%0           \n"
     "paddusb   %%xmm1,%%xmm0                   \n"
     "movdqa    %%xmm0,%%xmm2                   \n"
@@ -4893,25 +4186,20 @@ void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
     "punpckhwd %%xmm0,%%xmm0                   \n"
     "por       %%xmm5,%%xmm3                   \n"
     "por       %%xmm5,%%xmm0                   \n"
-    "sub       $0x10,%3                        \n"
-    "movdqa    %%xmm1," MEMACCESS(2) "         \n"
-    "movdqa    %%xmm2," MEMACCESS2(0x10,2) "   \n"
-    "movdqa    %%xmm3," MEMACCESS2(0x20,2) "   \n"
-    "movdqa    %%xmm0," MEMACCESS2(0x30,2) "   \n"
+    "movdqu    %%xmm1," MEMACCESS(2) "         \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x10,2) "   \n"
+    "movdqu    %%xmm3," MEMACCESS2(0x20,2) "   \n"
+    "movdqu    %%xmm0," MEMACCESS2(0x30,2) "   \n"
     "lea       " MEMLEA(0x40,2) ",%2           \n"
+    "sub       $0x10,%3                        \n"
     "jg        1b                              \n"
   : "+r"(src_sobelx),  // %0
     "+r"(src_sobely),  // %1
     "+r"(dst_argb),    // %2
     "+r"(width)        // %3
   :
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
   );
 }
 #endif  // HAS_SOBELROW_SSE2
@@ -4928,26 +4216,21 @@ void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
     // 8 pixel loop.
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    MEMOPREG(movdqa,0x00,0,1,1,xmm1)           //  movdqa    (%0,%1,1),%%xmm1
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,1,1,xmm1)           //  movdqu    (%0,%1,1),%%xmm1
     "lea       " MEMLEA(0x10,0) ",%0           \n"
     "paddusb   %%xmm1,%%xmm0                   \n"
-    "sub       $0x10,%3                        \n"
-    "movdqa    %%xmm0," MEMACCESS(2) "         \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
     "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x10,%3                        \n"
     "jg        1b                              \n"
   : "+r"(src_sobelx),  // %0
     "+r"(src_sobely),  // %1
     "+r"(dst_y),       // %2
     "+r"(width)        // %3
   :
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1"
   );
 }
 #endif  // HAS_SOBELTOPLANEROW_SSE2
@@ -4967,8 +4250,8 @@ void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
     // 8 pixel loop.
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    MEMOPREG(movdqa,0x00,0,1,1,xmm1)           //  movdqa    (%0,%1,1),%%xmm1
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,1,1,xmm1)           //  movdqu    (%0,%1,1),%%xmm1
     "lea       " MEMLEA(0x10,0) ",%0           \n"
     "movdqa    %%xmm0,%%xmm2                   \n"
     "paddusb   %%xmm1,%%xmm2                   \n"
@@ -4984,25 +4267,20 @@ void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
     "movdqa    %%xmm1,%%xmm7                   \n"
     "punpcklwd %%xmm0,%%xmm7                   \n"
     "punpckhwd %%xmm0,%%xmm1                   \n"
-    "sub       $0x10,%3                        \n"
-    "movdqa    %%xmm6," MEMACCESS(2) "         \n"
-    "movdqa    %%xmm4," MEMACCESS2(0x10,2) "   \n"
-    "movdqa    %%xmm7," MEMACCESS2(0x20,2) "   \n"
-    "movdqa    %%xmm1," MEMACCESS2(0x30,2) "   \n"
+    "movdqu    %%xmm6," MEMACCESS(2) "         \n"
+    "movdqu    %%xmm4," MEMACCESS2(0x10,2) "   \n"
+    "movdqu    %%xmm7," MEMACCESS2(0x20,2) "   \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x30,2) "   \n"
     "lea       " MEMLEA(0x40,2) ",%2           \n"
+    "sub       $0x10,%3                        \n"
     "jg        1b                              \n"
   : "+r"(src_sobelx),  // %0
     "+r"(src_sobely),  // %1
     "+r"(dst_argb),    // %2
     "+r"(width)        // %3
   :
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
   );
 }
 #endif  // HAS_SOBELXYROW_SSE2
@@ -5035,22 +4313,22 @@ void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
     "punpcklwd %%xmm1,%%xmm4                   \n"
     "punpckhwd %%xmm1,%%xmm5                   \n"
     "paddd     %%xmm2,%%xmm0                   \n"
-    "movdqa    " MEMACCESS(2) ",%%xmm2         \n"
+    "movdqu    " MEMACCESS(2) ",%%xmm2         \n"
     "paddd     %%xmm0,%%xmm2                   \n"
     "paddd     %%xmm3,%%xmm0                   \n"
-    "movdqa    " MEMACCESS2(0x10,2) ",%%xmm3   \n"
+    "movdqu    " MEMACCESS2(0x10,2) ",%%xmm3   \n"
     "paddd     %%xmm0,%%xmm3                   \n"
     "paddd     %%xmm4,%%xmm0                   \n"
-    "movdqa    " MEMACCESS2(0x20,2) ",%%xmm4   \n"
+    "movdqu    " MEMACCESS2(0x20,2) ",%%xmm4   \n"
     "paddd     %%xmm0,%%xmm4                   \n"
     "paddd     %%xmm5,%%xmm0                   \n"
-    "movdqa    " MEMACCESS2(0x30,2) ",%%xmm5   \n"
+    "movdqu    " MEMACCESS2(0x30,2) ",%%xmm5   \n"
     "lea       " MEMLEA(0x40,2) ",%2           \n"
     "paddd     %%xmm0,%%xmm5                   \n"
-    "movdqa    %%xmm2," MEMACCESS(1) "         \n"
-    "movdqa    %%xmm3," MEMACCESS2(0x10,1) "   \n"
-    "movdqa    %%xmm4," MEMACCESS2(0x20,1) "   \n"
-    "movdqa    %%xmm5," MEMACCESS2(0x30,1) "   \n"
+    "movdqu    %%xmm2," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm3," MEMACCESS2(0x10,1) "   \n"
+    "movdqu    %%xmm4," MEMACCESS2(0x20,1) "   \n"
+    "movdqu    %%xmm5," MEMACCESS2(0x30,1) "   \n"
     "lea       " MEMLEA(0x40,1) ",%1           \n"
     "sub       $0x4,%3                         \n"
     "jge       40b                             \n"
@@ -5082,9 +4360,7 @@ void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
     "+r"(width)  // %3
   :
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
   );
 }
 #endif  // HAS_COMPUTECUMULATIVESUMROW_SSE2
@@ -5115,11 +4391,10 @@ void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
   // 4 pixel small loop                        \n"
     LABELALIGN
   "4:                                         \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
-    BUNDLEALIGN
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
     MEMOPREG(psubd,0x00,0,4,4,xmm0)            // psubd    0x00(%0,%4,4),%%xmm0
     MEMOPREG(psubd,0x10,0,4,4,xmm1)            // psubd    0x10(%0,%4,4),%%xmm1
     MEMOPREG(psubd,0x20,0,4,4,xmm2)            // psubd    0x20(%0,%4,4),%%xmm2
@@ -5129,7 +4404,6 @@ void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
     "psubd     " MEMACCESS2(0x10,1) ",%%xmm1   \n"
     "psubd     " MEMACCESS2(0x20,1) ",%%xmm2   \n"
     "psubd     " MEMACCESS2(0x30,1) ",%%xmm3   \n"
-    BUNDLEALIGN
     MEMOPREG(paddd,0x00,1,4,4,xmm0)            // paddd    0x00(%1,%4,4),%%xmm0
     MEMOPREG(paddd,0x10,1,4,4,xmm1)            // paddd    0x10(%1,%4,4),%%xmm1
     MEMOPREG(paddd,0x20,1,4,4,xmm2)            // paddd    0x20(%1,%4,4),%%xmm2
@@ -5149,11 +4423,10 @@ void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
   // 4 pixel loop                              \n"
     LABELALIGN
   "40:                                         \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
-    BUNDLEALIGN
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
     MEMOPREG(psubd,0x00,0,4,4,xmm0)            // psubd    0x00(%0,%4,4),%%xmm0
     MEMOPREG(psubd,0x10,0,4,4,xmm1)            // psubd    0x10(%0,%4,4),%%xmm1
     MEMOPREG(psubd,0x20,0,4,4,xmm2)            // psubd    0x20(%0,%4,4),%%xmm2
@@ -5163,7 +4436,6 @@ void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
     "psubd     " MEMACCESS2(0x10,1) ",%%xmm1   \n"
     "psubd     " MEMACCESS2(0x20,1) ",%%xmm2   \n"
     "psubd     " MEMACCESS2(0x30,1) ",%%xmm3   \n"
-    BUNDLEALIGN
     MEMOPREG(paddd,0x00,1,4,4,xmm0)            // paddd    0x00(%1,%4,4),%%xmm0
     MEMOPREG(paddd,0x10,1,4,4,xmm1)            // paddd    0x10(%1,%4,4),%%xmm1
     MEMOPREG(paddd,0x20,1,4,4,xmm2)            // paddd    0x20(%1,%4,4),%%xmm2
@@ -5196,11 +4468,10 @@ void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
   // 1 pixel loop                              \n"
     LABELALIGN
   "10:                                         \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
     MEMOPREG(psubd,0x00,0,4,4,xmm0)            // psubd    0x00(%0,%4,4),%%xmm0
     "lea       " MEMLEA(0x10,0) ",%0           \n"
     "psubd     " MEMACCESS(1) ",%%xmm0         \n"
-    BUNDLEALIGN
     MEMOPREG(paddd,0x00,1,4,4,xmm0)            // paddd    0x00(%1,%4,4),%%xmm0
     "lea       " MEMLEA(0x10,1) ",%1           \n"
     "cvtdq2ps  %%xmm0,%%xmm0                   \n"
@@ -5219,13 +4490,8 @@ void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
     "+rm"(count)    // %3
   : "r"((intptr_t)(width)),  // %4
     "rm"(area)     // %5
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
   );
 }
 #endif  // HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
@@ -5268,7 +4534,6 @@ void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
     "pshufd    $0x39,%%xmm0,%%xmm0             \n"
     "movd      %%xmm0,%k5                      \n"
     "pshufd    $0x39,%%xmm0,%%xmm0             \n"
-    BUNDLEALIGN
     MEMOPREG(movd,0x00,0,1,1,xmm1)             //  movd      (%0,%1,1),%%xmm1
     MEMOPREG(movd,0x00,0,5,1,xmm6)             //  movd      (%0,%5,1),%%xmm6
     "punpckldq %%xmm6,%%xmm1                   \n"
@@ -5277,14 +4542,13 @@ void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
     "movd      %%xmm0,%k1                      \n"
     "pshufd    $0x39,%%xmm0,%%xmm0             \n"
     "movd      %%xmm0,%k5                      \n"
-    BUNDLEALIGN
     MEMOPREG(movd,0x00,0,1,1,xmm0)             //  movd      (%0,%1,1),%%xmm0
     MEMOPREG(movd,0x00,0,5,1,xmm6)             //  movd      (%0,%5,1),%%xmm6
     "punpckldq %%xmm6,%%xmm0                   \n"
     "addps     %%xmm4,%%xmm3                   \n"
-    "sub       $0x4,%4                         \n"
     "movq      %%xmm0," MEMACCESS2(0x08,2) "   \n"
     "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%4                         \n"
     "jge       40b                             \n"
 
   "49:                                         \n"
@@ -5299,11 +4563,10 @@ void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
     "pmaddwd   %%xmm5,%%xmm0                   \n"
     "addps     %%xmm7,%%xmm2                   \n"
     "movd      %%xmm0,%k1                      \n"
-    BUNDLEALIGN
     MEMOPREG(movd,0x00,0,1,1,xmm0)             //  movd      (%0,%1,1),%%xmm0
-    "sub       $0x1,%4                         \n"
     "movd      %%xmm0," MEMACCESS(2) "         \n"
     "lea       " MEMLEA(0x04,2) ",%2           \n"
+    "sub       $0x1,%4                         \n"
     "jge       10b                             \n"
   "19:                                         \n"
   : "+r"(src_argb),  // %0
@@ -5313,13 +4576,8 @@ void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
     "+rm"(width),    // %4
     "+r"(temp)   // %5
   :
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
   );
 }
 #endif  // HAS_ARGBAFFINEROW_SSE2
@@ -5352,8 +4610,8 @@ void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
     // General purpose row blend.
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
-    MEMOPREG(movdqa,0x00,1,4,1,xmm2)
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm2)
     "movdqa    %%xmm0,%%xmm1                   \n"
     "punpcklbw %%xmm2,%%xmm0                   \n"
     "punpckhbw %%xmm2,%%xmm1                   \n"
@@ -5362,61 +4620,57 @@ void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
     "psrlw     $0x7,%%xmm0                     \n"
     "psrlw     $0x7,%%xmm1                     \n"
     "packuswb  %%xmm1,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    BUNDLEALIGN
-    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
     "jg        1b                              \n"
     "jmp       99f                             \n"
 
     // Blend 25 / 75.
     LABELALIGN
   "25:                                         \n"
-    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
-    MEMOPREG(movdqa,0x00,1,4,1,xmm1)
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm1)
     "pavgb     %%xmm1,%%xmm0                   \n"
     "pavgb     %%xmm1,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    BUNDLEALIGN
-    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
     "jg        25b                             \n"
     "jmp       99f                             \n"
 
     // Blend 50 / 50.
     LABELALIGN
   "50:                                         \n"
-    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
-    MEMOPREG(movdqa,0x00,1,4,1,xmm1)
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm1)
     "pavgb     %%xmm1,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    BUNDLEALIGN
-    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
     "jg        50b                             \n"
     "jmp       99f                             \n"
 
     // Blend 75 / 25.
     LABELALIGN
   "75:                                         \n"
-    "movdqa    " MEMACCESS(1) ",%%xmm1         \n"
-    MEMOPREG(movdqa,0x00,1,4,1,xmm0)
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm0)
     "pavgb     %%xmm1,%%xmm0                   \n"
     "pavgb     %%xmm1,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    BUNDLEALIGN
-    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
     "jg        75b                             \n"
     "jmp       99f                             \n"
 
     // Blend 100 / 0 - Copy row unchanged.
     LABELALIGN
   "100:                                        \n"
-    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
-    "sub       $0x10,%2                        \n"
-    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
     "jg        100b                            \n"
 
   "99:                                         \n"
@@ -5425,147 +4679,22 @@ void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
     "+r"(dst_width),  // %2
     "+r"(source_y_fraction)  // %3
   : "r"((intptr_t)(src_stride))  // %4
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm5"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm5"
   );
 }
 #endif  // HAS_INTERPOLATEROW_SSSE3
 
-#ifdef HAS_INTERPOLATEROW_SSE2
-// Bilinear filter 16x2 -> 16x1
-void InterpolateRow_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+#ifdef HAS_INTERPOLATEROW_AVX2
+// Bilinear filter 32x2 -> 32x1
+void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
                          ptrdiff_t src_stride, int dst_width,
                          int source_y_fraction) {
   asm volatile (
-    "sub       %1,%0                           \n"
     "shr       %3                              \n"
     "cmp       $0x0,%3                         \n"
     "je        100f                            \n"
-    "cmp       $0x20,%3                        \n"
-    "je        75f                             \n"
-    "cmp       $0x40,%3                        \n"
-    "je        50f                             \n"
-    "cmp       $0x60,%3                        \n"
-    "je        25f                             \n"
-
-    "movd      %3,%%xmm0                       \n"
-    "neg       %3                              \n"
-    "add       $0x80,%3                        \n"
-    "movd      %3,%%xmm5                       \n"
-    "punpcklbw %%xmm0,%%xmm5                   \n"
-    "punpcklwd %%xmm5,%%xmm5                   \n"
-    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
-    "pxor      %%xmm4,%%xmm4                   \n"
-
-    // General purpose row blend.
-    LABELALIGN
-  "1:                                          \n"
-    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
-    MEMOPREG(movdqa,0x00,1,4,1,xmm2)           //  movdqa    (%1,%4,1),%%xmm2
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "movdqa    %%xmm2,%%xmm3                   \n"
-    "punpcklbw %%xmm4,%%xmm2                   \n"
-    "punpckhbw %%xmm4,%%xmm3                   \n"
-    "punpcklbw %%xmm4,%%xmm0                   \n"
-    "punpckhbw %%xmm4,%%xmm1                   \n"
-    "psubw     %%xmm0,%%xmm2                   \n"
-    "psubw     %%xmm1,%%xmm3                   \n"
-    "paddw     %%xmm2,%%xmm2                   \n"
-    "paddw     %%xmm3,%%xmm3                   \n"
-    "pmulhw    %%xmm5,%%xmm2                   \n"
-    "pmulhw    %%xmm5,%%xmm3                   \n"
-    "paddw     %%xmm2,%%xmm0                   \n"
-    "paddw     %%xmm3,%%xmm1                   \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    BUNDLEALIGN
-    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)           //  movdqa    %%xmm0,(%1,%0,1)
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "jg        1b                              \n"
-    "jmp       99f                             \n"
-
-    // Blend 25 / 75.
-    LABELALIGN
-  "25:                                         \n"
-    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
-    MEMOPREG(movdqa,0x00,1,4,1,xmm1)           //  movdqa    (%1,%4,1),%%xmm1
-    "pavgb     %%xmm1,%%xmm0                   \n"
-    "pavgb     %%xmm1,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    BUNDLEALIGN
-    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)           //  movdqa    %%xmm0,(%1,%0,1)
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "jg        25b                             \n"
-    "jmp       99f                             \n"
-
-    // Blend 50 / 50.
-    LABELALIGN
-  "50:                                         \n"
-    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
-    MEMOPREG(movdqa,0x00,1,4,1,xmm1)           //  movdqa    (%1,%4,1),%%xmm1
-    "pavgb     %%xmm1,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    BUNDLEALIGN
-    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)           //  movdqa    %%xmm0,(%1,%0,1)
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "jg        50b                             \n"
-    "jmp       99f                             \n"
-
-    // Blend 75 / 25.
-    LABELALIGN
-  "75:                                         \n"
-    "movdqa    " MEMACCESS(1) ",%%xmm1         \n"
-    MEMOPREG(movdqa,0x00,1,4,1,xmm0)           //  movdqa    (%1,%4,1),%%xmm0
-    "pavgb     %%xmm1,%%xmm0                   \n"
-    "pavgb     %%xmm1,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    BUNDLEALIGN
-    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)           //  movdqa    %%xmm0,(%1,%0,1)
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "jg        75b                             \n"
-    "jmp       99f                             \n"
-
-    // Blend 100 / 0 - Copy row unchanged.
-    LABELALIGN
-  "100:                                        \n"
-    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
-    "sub       $0x10,%2                        \n"
-    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)           //  movdqa    %%xmm0,(%1,%0,1)
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "jg        100b                            \n"
-
-  "99:                                         \n"
-  : "+r"(dst_ptr),    // %0
-    "+r"(src_ptr),    // %1
-    "+r"(dst_width),  // %2
-    "+r"(source_y_fraction)  // %3
-  : "r"((intptr_t)(src_stride))  // %4
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
-  );
-}
-#endif  // HAS_INTERPOLATEROW_SSE2
-
-#ifdef HAS_INTERPOLATEROW_SSSE3
-// Bilinear filter 16x2 -> 16x1
-void InterpolateRow_Unaligned_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
-                                    ptrdiff_t src_stride, int dst_width,
-                                    int source_y_fraction) {
-  asm volatile (
     "sub       %1,%0                           \n"
-    "shr       %3                              \n"
-    "cmp       $0x0,%3                         \n"
-    "je        100f                            \n"
     "cmp       $0x20,%3                        \n"
     "je        75f                             \n"
     "cmp       $0x40,%3                        \n"
@@ -5573,106 +4702,95 @@ void InterpolateRow_Unaligned_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
     "cmp       $0x60,%3                        \n"
     "je        25f                             \n"
 
-    "movd      %3,%%xmm0                       \n"
-    "neg       %3                              \n"
-    "add       $0x80,%3                        \n"
-    "movd      %3,%%xmm5                       \n"
-    "punpcklbw %%xmm0,%%xmm5                   \n"
-    "punpcklwd %%xmm5,%%xmm5                   \n"
-    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "vmovd      %3,%%xmm0                      \n"
+    "neg        %3                             \n"
+    "add        $0x80,%3                       \n"
+    "vmovd      %3,%%xmm5                      \n"
+    "vpunpcklbw %%xmm0,%%xmm5,%%xmm5           \n"
+    "vpunpcklwd %%xmm5,%%xmm5,%%xmm5           \n"
+    "vpxor      %%ymm0,%%ymm0,%%ymm0           \n"
+    "vpermd     %%ymm5,%%ymm0,%%ymm5           \n"
 
     // General purpose row blend.
     LABELALIGN
   "1:                                          \n"
-    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
-    MEMOPREG(movdqu,0x00,1,4,1,xmm2)
-    "movdqu    %%xmm0,%%xmm1                   \n"
-    "punpcklbw %%xmm2,%%xmm0                   \n"
-    "punpckhbw %%xmm2,%%xmm1                   \n"
-    "pmaddubsw %%xmm5,%%xmm0                   \n"
-    "pmaddubsw %%xmm5,%%xmm1                   \n"
-    "psrlw     $0x7,%%xmm0                     \n"
-    "psrlw     $0x7,%%xmm1                     \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    BUNDLEALIGN
-    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "vmovdqu    " MEMACCESS(1) ",%%ymm0        \n"
+    MEMOPREG(vmovdqu,0x00,1,4,1,ymm2)
+    "vpunpckhbw %%ymm2,%%ymm0,%%ymm1           \n"
+    "vpunpcklbw %%ymm2,%%ymm0,%%ymm0           \n"
+    "vpmaddubsw %%ymm5,%%ymm0,%%ymm0           \n"
+    "vpmaddubsw %%ymm5,%%ymm1,%%ymm1           \n"
+    "vpsrlw     $0x7,%%ymm0,%%ymm0             \n"
+    "vpsrlw     $0x7,%%ymm1,%%ymm1             \n"
+    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
+    MEMOPMEM(vmovdqu,ymm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
     "jg        1b                              \n"
     "jmp       99f                             \n"
 
     // Blend 25 / 75.
     LABELALIGN
   "25:                                         \n"
-    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
-    MEMOPREG(movdqu,0x00,1,4,1,xmm1)
-    "pavgb     %%xmm1,%%xmm0                   \n"
-    "pavgb     %%xmm1,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    BUNDLEALIGN
-    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "vmovdqu    " MEMACCESS(1) ",%%ymm0        \n"
+    MEMOPREG(vmovdqu,0x00,1,4,1,ymm1)
+    "vpavgb     %%ymm1,%%ymm0,%%ymm0           \n"
+    "vpavgb     %%ymm1,%%ymm0,%%ymm0           \n"
+    MEMOPMEM(vmovdqu,ymm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
     "jg        25b                             \n"
     "jmp       99f                             \n"
 
     // Blend 50 / 50.
     LABELALIGN
   "50:                                         \n"
-    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
-    MEMOPREG(movdqu,0x00,1,4,1,xmm1)
-    "pavgb     %%xmm1,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    BUNDLEALIGN
-    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "vmovdqu    " MEMACCESS(1) ",%%ymm0        \n"
+    VMEMOPREG(vpavgb,0x00,1,4,1,ymm0,ymm0)     // vpavgb (%1,%4,1),%%ymm0,%%ymm0
+    MEMOPMEM(vmovdqu,ymm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
     "jg        50b                             \n"
     "jmp       99f                             \n"
 
     // Blend 75 / 25.
     LABELALIGN
   "75:                                         \n"
-    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
-    MEMOPREG(movdqu,0x00,1,4,1,xmm0)
-    "pavgb     %%xmm1,%%xmm0                   \n"
-    "pavgb     %%xmm1,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    BUNDLEALIGN
-    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "vmovdqu    " MEMACCESS(1) ",%%ymm1        \n"
+    MEMOPREG(vmovdqu,0x00,1,4,1,ymm0)
+    "vpavgb     %%ymm1,%%ymm0,%%ymm0           \n"
+    "vpavgb     %%ymm1,%%ymm0,%%ymm0           \n"
+    MEMOPMEM(vmovdqu,ymm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
     "jg        75b                             \n"
     "jmp       99f                             \n"
 
     // Blend 100 / 0 - Copy row unchanged.
     LABELALIGN
   "100:                                        \n"
-    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
-    "sub       $0x10,%2                        \n"
-    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "jg        100b                            \n"
+    "rep movsb " MEMMOVESTRING(1,0) "          \n"
+    "jmp       999f                            \n"
 
   "99:                                         \n"
-  : "+r"(dst_ptr),    // %0
-    "+r"(src_ptr),    // %1
-    "+r"(dst_width),  // %2
+    "vzeroupper                                \n"
+  "999:                                        \n"
+  : "+D"(dst_ptr),    // %0
+    "+S"(src_ptr),    // %1
+    "+c"(dst_width),  // %2
     "+r"(source_y_fraction)  // %3
   : "r"((intptr_t)(src_stride))  // %4
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm5"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm5"
   );
 }
-#endif   // HAS_INTERPOLATEROW_SSSE3
+#endif  // HAS_INTERPOLATEROW_AVX2
 
 #ifdef HAS_INTERPOLATEROW_SSE2
 // Bilinear filter 16x2 -> 16x1
-void InterpolateRow_Unaligned_SSE2(uint8* dst_ptr, const uint8* src_ptr,
-                                   ptrdiff_t src_stride, int dst_width,
-                                   int source_y_fraction) {
+void InterpolateRow_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) {
   asm volatile (
     "sub       %1,%0                           \n"
     "shr       %3                              \n"
@@ -5699,8 +4817,8 @@ void InterpolateRow_Unaligned_SSE2(uint8* dst_ptr, const uint8* src_ptr,
   "1:                                          \n"
     "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
     MEMOPREG(movdqu,0x00,1,4,1,xmm2)           //  movdqu    (%1,%4,1),%%xmm2
-    "movdqu    %%xmm0,%%xmm1                   \n"
-    "movdqu    %%xmm2,%%xmm3                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
     "punpcklbw %%xmm4,%%xmm2                   \n"
     "punpckhbw %%xmm4,%%xmm3                   \n"
     "punpcklbw %%xmm4,%%xmm0                   \n"
@@ -5714,10 +4832,9 @@ void InterpolateRow_Unaligned_SSE2(uint8* dst_ptr, const uint8* src_ptr,
     "paddw     %%xmm2,%%xmm0                   \n"
     "paddw     %%xmm3,%%xmm1                   \n"
     "packuswb  %%xmm1,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    BUNDLEALIGN
     MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
     "jg        1b                              \n"
     "jmp       99f                             \n"
 
@@ -5728,10 +4845,9 @@ void InterpolateRow_Unaligned_SSE2(uint8* dst_ptr, const uint8* src_ptr,
     MEMOPREG(movdqu,0x00,1,4,1,xmm1)           //  movdqu    (%1,%4,1),%%xmm1
     "pavgb     %%xmm1,%%xmm0                   \n"
     "pavgb     %%xmm1,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    BUNDLEALIGN
     MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
     "jg        25b                             \n"
     "jmp       99f                             \n"
 
@@ -5741,10 +4857,9 @@ void InterpolateRow_Unaligned_SSE2(uint8* dst_ptr, const uint8* src_ptr,
     "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
     MEMOPREG(movdqu,0x00,1,4,1,xmm1)           //  movdqu    (%1,%4,1),%%xmm1
     "pavgb     %%xmm1,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    BUNDLEALIGN
     MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
     "jg        50b                             \n"
     "jmp       99f                             \n"
 
@@ -5755,10 +4870,9 @@ void InterpolateRow_Unaligned_SSE2(uint8* dst_ptr, const uint8* src_ptr,
     MEMOPREG(movdqu,0x00,1,4,1,xmm0)           //  movdqu    (%1,%4,1),%%xmm0
     "pavgb     %%xmm1,%%xmm0                   \n"
     "pavgb     %%xmm1,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    BUNDLEALIGN
     MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
     "jg        75b                             \n"
     "jmp       99f                             \n"
 
@@ -5766,9 +4880,9 @@ void InterpolateRow_Unaligned_SSE2(uint8* dst_ptr, const uint8* src_ptr,
     LABELALIGN
   "100:                                        \n"
     "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
-    "sub       $0x10,%2                        \n"
     MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
     "jg        100b                            \n"
 
   "99:                                         \n"
@@ -5777,73 +4891,12 @@ void InterpolateRow_Unaligned_SSE2(uint8* dst_ptr, const uint8* src_ptr,
     "+r"(dst_width),  // %2
     "+r"(source_y_fraction)  // %3
   : "r"((intptr_t)(src_stride))  // %4
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
 }
 #endif  // HAS_INTERPOLATEROW_SSE2
 
-#ifdef HAS_HALFROW_SSE2
-void HalfRow_SSE2(const uint8* src_uv, int src_uv_stride,
-                  uint8* dst_uv, int pix) {
-  asm volatile (
-    "sub       %0,%1                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    MEMOPREG(pavgb,0x00,0,3,1,xmm0)            //  pavgb     (%0,%3),%%xmm0
-    "sub       $0x10,%2                        \n"
-    MEMOPMEM(movdqa,xmm0,0x00,0,1,1)           //  movdqa    %%xmm0,(%0,%1)
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "jg        1b                              \n"
-  : "+r"(src_uv),  // %0
-    "+r"(dst_uv),  // %1
-    "+r"(pix)      // %2
-  : "r"((intptr_t)(src_uv_stride))  // %3
-  : "memory", "cc"
-#if defined(__SSE2__)
-      , "xmm0"
-#endif
-  );
-}
-#endif  // HAS_HALFROW_SSE2
-
-#ifdef HAS_ARGBTOBAYERROW_SSSE3
-void ARGBToBayerRow_SSSE3(const uint8* src_argb, uint8* dst_bayer,
-                          uint32 selector, int pix) {
-  asm volatile (
-    // NaCL caveat - assumes movd is from GPR
-    "movd      %3,%%xmm5                       \n"
-    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "pshufb    %%xmm5,%%xmm0                   \n"
-    "pshufb    %%xmm5,%%xmm1                   \n"
-    "punpckldq %%xmm1,%%xmm0                   \n"
-    "sub       $0x8,%2                         \n"
-    "movq      %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_bayer), // %1
-    "+r"(pix)        // %2
-  : "g"(selector)    // %3
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm5"
-#endif
-  );
-}
-#endif  // HAS_ARGBTOBAYERROW_SSSE3
-
 #ifdef HAS_ARGBTOBAYERGGROW_SSE2
 void ARGBToBayerGGRow_SSE2(const uint8* src_argb, uint8* dst_bayer,
                            uint32 selector, int pix) {
@@ -5852,8 +4905,8 @@ void ARGBToBayerGGRow_SSE2(const uint8* src_argb, uint8* dst_bayer,
     "psrld     $0x18,%%xmm5                    \n"
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
     "lea       " MEMLEA(0x20,0) ",%0           \n"
     "psrld     $0x8,%%xmm0                     \n"
     "psrld     $0x8,%%xmm1                     \n"
@@ -5861,18 +4914,16 @@ void ARGBToBayerGGRow_SSE2(const uint8* src_argb, uint8* dst_bayer,
     "pand      %%xmm5,%%xmm1                   \n"
     "packssdw  %%xmm1,%%xmm0                   \n"
     "packuswb  %%xmm1,%%xmm0                   \n"
-    "sub       $0x8,%2                         \n"
     "movq      %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x8,%2                         \n"
     "jg        1b                              \n"
   : "+r"(src_argb),  // %0
     "+r"(dst_bayer), // %1
     "+r"(pix)        // %2
   :
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1", "xmm5"
-#endif
   );
 }
 #endif  // HAS_ARGBTOBAYERGGROW_SSE2
@@ -5882,34 +4933,7 @@ void ARGBToBayerGGRow_SSE2(const uint8* src_argb, uint8* dst_bayer,
 void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
                           const uint8* shuffler, int pix) {
   asm volatile (
-    "movdqa    " MEMACCESS(3) ",%%xmm5         \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "pshufb    %%xmm5,%%xmm0                   \n"
-    "pshufb    %%xmm5,%%xmm1                   \n"
-    "sub       $0x8,%2                         \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
-    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
-    "lea       " MEMLEA(0x20,1) ",%1           \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_argb),  // %1
-    "+r"(pix)        // %2
-  : "r"(shuffler)    // %3
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm5"
-#endif
-  );
-}
-
-void ARGBShuffleRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_argb,
-                                    const uint8* shuffler, int pix) {
-  asm volatile (
-    "movdqa    " MEMACCESS(3) ",%%xmm5         \n"
+    "movdqu    " MEMACCESS(3) ",%%xmm5         \n"
     LABELALIGN
   "1:                                          \n"
     "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
@@ -5917,19 +4941,17 @@ void ARGBShuffleRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_argb,
     "lea       " MEMLEA(0x20,0) ",%0           \n"
     "pshufb    %%xmm5,%%xmm0                   \n"
     "pshufb    %%xmm5,%%xmm1                   \n"
-    "sub       $0x8,%2                         \n"
     "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
     "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
     "jg        1b                              \n"
   : "+r"(src_argb),  // %0
     "+r"(dst_argb),  // %1
     "+r"(pix)        // %2
   : "r"(shuffler)    // %3
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1", "xmm5"
-#endif
   );
 }
 #endif  // HAS_ARGBSHUFFLEROW_SSSE3
@@ -5947,19 +4969,18 @@ void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb,
     "lea       " MEMLEA(0x40,0) ",%0           \n"
     "vpshufb   %%ymm5,%%ymm0,%%ymm0            \n"
     "vpshufb   %%ymm5,%%ymm1,%%ymm1            \n"
-    "sub       $0x10,%2                        \n"
     "vmovdqu   %%ymm0," MEMACCESS(1) "         \n"
     "vmovdqu   %%ymm1," MEMACCESS2(0x20,1) "   \n"
     "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
     "jg        1b                              \n"
+    "vzeroupper                                \n"
   : "+r"(src_argb),  // %0
     "+r"(dst_argb),  // %1
     "+r"(pix)        // %2
   : "r"(shuffler)    // %3
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1", "xmm5"
-#endif
   );
 }
 #endif  // HAS_ARGBSHUFFLEROW_AVX2
@@ -5989,7 +5010,6 @@ void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
     "movzb     " MEMACCESS2(0x1,4) ",%2        \n"
     MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
     "mov       %b2," MEMACCESS2(0x1,1) "       \n"
-    BUNDLEALIGN
     "movzb     " MEMACCESS2(0x2,4) ",%2        \n"
     MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
     "mov       %b2," MEMACCESS2(0x2,1) "       \n"
@@ -6014,9 +5034,9 @@ void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
     "pshufhw   $0x1b,%%xmm1,%%xmm1             \n"
     "pshuflw   $0x1b,%%xmm1,%%xmm1             \n"
     "packuswb  %%xmm1,%%xmm0                   \n"
-    "sub       $0x4,%3                         \n"
     "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%3                         \n"
     "jg        123b                            \n"
     "jmp       99f                             \n"
 
@@ -6032,9 +5052,9 @@ void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
     "pshufhw   $0x39,%%xmm1,%%xmm1             \n"
     "pshuflw   $0x39,%%xmm1,%%xmm1             \n"
     "packuswb  %%xmm1,%%xmm0                   \n"
-    "sub       $0x4,%3                         \n"
     "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%3                         \n"
     "jg        321b                            \n"
     "jmp       99f                             \n"
 
@@ -6050,9 +5070,9 @@ void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
     "pshufhw   $0x93,%%xmm1,%%xmm1             \n"
     "pshuflw   $0x93,%%xmm1,%%xmm1             \n"
     "packuswb  %%xmm1,%%xmm0                   \n"
-    "sub       $0x4,%3                         \n"
     "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%3                         \n"
     "jg        2103b                           \n"
     "jmp       99f                             \n"
 
@@ -6068,9 +5088,9 @@ void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
     "pshufhw   $0xc6,%%xmm1,%%xmm1             \n"
     "pshuflw   $0xc6,%%xmm1,%%xmm1             \n"
     "packuswb  %%xmm1,%%xmm0                   \n"
-    "sub       $0x4,%3                         \n"
     "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%3                         \n"
     "jg        3012b                           \n"
 
   "99:                                         \n"
@@ -6079,13 +5099,8 @@ void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
     "+d"(pixel_temp),  // %2
     "+r"(pix)         // %3
   : "r"(shuffler)      // %4
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm5"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm5"
   );
 }
 #endif  // HAS_ARGBSHUFFLEROW_SSE2
@@ -6119,13 +5134,8 @@ void I422ToYUY2Row_SSE2(const uint8* src_y,
       "+r"(dst_frame),  // %3
       "+rm"(width)  // %4
     :
-    : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3"
-#endif
+    : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3"
   );
 }
 #endif  // HAS_I422TOYUY2ROW_SSE2
@@ -6159,13 +5169,8 @@ void I422ToUYVYRow_SSE2(const uint8* src_y,
       "+r"(dst_frame),  // %3
       "+rm"(width)  // %4
     :
-    : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3"
-#endif
+    : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3"
   );
 }
 #endif  // HAS_I422TOUYVYROW_SSE2
@@ -6212,18 +5217,16 @@ void ARGBPolynomialRow_SSE2(const uint8* src_argb,
     "cvttps2dq %%xmm4,%%xmm4                   \n"
     "packuswb  %%xmm4,%%xmm0                   \n"
     "packuswb  %%xmm0,%%xmm0                   \n"
-    "sub       $0x2,%2                         \n"
     "movq      %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x2,%2                         \n"
     "jg        1b                              \n"
   : "+r"(src_argb),  // %0
     "+r"(dst_argb),  // %1
     "+r"(width)      // %2
   : "r"(poly)        // %3
   : "memory", "cc"
-#if defined(__SSE2__)
     , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
-#endif
   );
 }
 #endif  // HAS_ARGBPOLYNOMIALROW_SSE2
@@ -6253,20 +5256,17 @@ void ARGBPolynomialRow_AVX2(const uint8* src_argb,
     "vpackusdw   %%ymm0,%%ymm0,%%ymm0          \n"
     "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
     "vpackuswb   %%xmm0,%%xmm0,%%xmm0          \n"
-    "sub         $0x2,%2                       \n"
     "vmovq       %%xmm0," MEMACCESS(1) "       \n"
     "lea         " MEMLEA(0x8,1) ",%1          \n"
+    "sub         $0x2,%2                       \n"
     "jg          1b                            \n"
     "vzeroupper                                \n"
   : "+r"(src_argb),  // %0
     "+r"(dst_argb),  // %1
     "+r"(width)      // %2
   : "r"(poly)        // %3
-  : "memory", "cc"
-#if defined(__SSE2__)
-// TODO(fbarchard): declare ymm usage when applicable.
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-#endif
+  : "memory", "cc",
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
   );
 }
 #endif  // HAS_ARGBPOLYNOMIALROW_AVX2
@@ -6376,7 +5376,6 @@ void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
     "movzb     " MEMACCESS2(0x4,2) ",%0        \n"
     MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
     "mov       %b0," MEMACCESS2(0x4,3) "       \n"
-    BUNDLEALIGN
     "movzb     " MEMACCESS2(0x5,2) ",%0        \n"
     MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
     "mov       %b0," MEMACCESS2(0x5,3) "       \n"
@@ -6416,9 +5415,9 @@ void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
     "mov       %b0," MEMACCESS2(0xe,3) "       \n"
     "movzb     " MEMACCESS2(0xf,2) ",%0        \n"
     "mov       %b0," MEMACCESS2(0xf,3) "       \n"
-    "sub       $0x4,%4                         \n"
     "lea       " MEMLEA(0x10,2) ",%2           \n"
     "lea       " MEMLEA(0x10,3) ",%3           \n"
+    "sub       $0x4,%4                         \n"
     "jg        1b                              \n"
   : "+d"(pixel_temp),  // %0
     "+a"(table_temp),  // %1
@@ -6427,10 +5426,7 @@ void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
     "+rm"(width)       // %4
   : "r"(luma),         // %5
     "rm"(lumacoeff)    // %6
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm3", "xmm4", "xmm5"
-#endif
+  : "memory", "cc", "xmm0", "xmm3", "xmm4", "xmm5"
   );
 }
 #endif  // HAS_ARGBLUMACOLORTABLEROW_SSSE3
diff --git a/third_party/libyuv/source/row_win.cc b/third_party/libyuv/source/row_win.cc
index d79c35396..5c06b6078 100644
--- a/third_party/libyuv/source/row_win.cc
+++ b/third_party/libyuv/source/row_win.cc
@@ -24,55 +24,63 @@ extern "C" {
 #if !defined(LIBYUV_DISABLE_X86) && defined(_MSC_VER) && \
     (defined(_M_IX86) || defined(_M_X64))
 
-#define YG 74  /* (int8)(1.164 * 64 + 0.5) */
-
-#define UB 127  /* min(127,(int8)(2.018 * 64)) */
-#define UG -25  /* (int8)(-0.391 * 64 - 0.5) */
-#define UR 0
-
-#define VB 0
-#define VG -52  /* (int8)(-0.813 * 64 - 0.5) */
-#define VR 102  /* (int8)(1.596 * 64 + 0.5) */
-
-// Bias
-#define BB UB * 128 + VB * 128
-#define BG UG * 128 + VG * 128
-#define BR UR * 128 + VR * 128
-
-static const vec8 kUVToB = {
-  UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB
-};
-
-static const vec8 kUVToR = {
-  UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR
-};
-
-static const vec8 kUVToG = {
-  UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG
+// YUV to RGB conversion constants.
+// Y contribution to R,G,B.  Scale and bias.
+#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */
+#define YGB 1160 /* 1.164 * 64 * 16 - adjusted for even error distribution */
+
+// U and V contributions to R,G,B.
+#define UB -128 /* -min(128, round(2.018 * 64)) */
+#define UG 25 /* -round(-0.391 * 64) */
+#define VG 52 /* -round(-0.813 * 64) */
+#define VR -102 /* -round(1.596 * 64) */
+
+// Bias values to subtract 16 from Y and 128 from U and V.
+#define BB (UB * 128            - YGB)
+#define BG (UG * 128 + VG * 128 - YGB)
+#define BR            (VR * 128 - YGB)
+
+struct YuvConstants {
+  lvec8 kUVToB;     // 0
+  lvec8 kUVToG;     // 32
+  lvec8 kUVToR;     // 64
+  lvec16 kUVBiasB;  // 96
+  lvec16 kUVBiasG;  // 128
+  lvec16 kUVBiasR;  // 160
+  lvec16 kYToRgb;   // 192
 };
 
-static const vec8 kVUToB = {
-  VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB,
+// BT601 constants for YUV to RGB.
+static YuvConstants SIMD_ALIGNED(kYuvConstants) = {
+  { UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0,
+    UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0 },
+  { UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG,
+    UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG },
+  { 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR,
+    0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR },
+  { BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB },
+  { BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG },
+  { BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR },
+  { YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG }
 };
 
-static const vec8 kVUToR = {
-  VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR,
+// BT601 constants for NV21 where chroma plane is VU instead of UV.
+static YuvConstants SIMD_ALIGNED(kYvuConstants) = {
+  { 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB,
+    0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB },
+  { VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG,
+    VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG },
+  { VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0,
+    VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0 },
+  { BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB },
+  { BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG },
+  { BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR },
+  { YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG }
 };
 
-static const vec8 kVUToG = {
-  VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG,
-};
-
-static const vec16 kYToRgb = { YG, YG, YG, YG, YG, YG, YG, YG };
-static const vec16 kYSub16 = { 16, 16, 16, 16, 16, 16, 16, 16 };
-static const vec16 kUVBiasB = { BB, BB, BB, BB, BB, BB, BB, BB };
-static const vec16 kUVBiasG = { BG, BG, BG, BG, BG, BG, BG, BG };
-static const vec16 kUVBiasR = { BR, BR, BR, BR, BR, BR, BR, BR };
-
 // 64 bit
 #if defined(_M_X64)
 
-// Aligned destination version.
 __declspec(align(16))
 void I422ToARGBRow_SSSE3(const uint8* y_buf,
                          const uint8* u_buf,
@@ -81,7 +89,6 @@ void I422ToARGBRow_SSSE3(const uint8* y_buf,
                          int width) {
   __m128i xmm0, xmm1, xmm2, xmm3;
   const __m128i xmm5 = _mm_set1_epi8(-1);
-  const __m128i xmm4 = _mm_setzero_si128();
   const ptrdiff_t offset = (uint8*)v_buf - (uint8*)u_buf;
 
   while (width > 0) {
@@ -89,18 +96,17 @@ void I422ToARGBRow_SSSE3(const uint8* y_buf,
     xmm1 = _mm_cvtsi32_si128(*(uint32*)(u_buf + offset));
     xmm0 = _mm_unpacklo_epi8(xmm0, xmm1);
     xmm0 = _mm_unpacklo_epi16(xmm0, xmm0);
-    xmm1 = _mm_load_si128(&xmm0);
-    xmm2 = _mm_load_si128(&xmm0);
-    xmm0 = _mm_maddubs_epi16(xmm0, *(__m128i*)kUVToB);
-    xmm1 = _mm_maddubs_epi16(xmm1, *(__m128i*)kUVToG);
-    xmm2 = _mm_maddubs_epi16(xmm2, *(__m128i*)kUVToR);
-    xmm0 = _mm_sub_epi16(xmm0, *(__m128i*)kUVBiasB);
-    xmm1 = _mm_sub_epi16(xmm1, *(__m128i*)kUVBiasG);
-    xmm2 = _mm_sub_epi16(xmm2, *(__m128i*)kUVBiasR);
+    xmm1 = _mm_loadu_si128(&xmm0);
+    xmm2 = _mm_loadu_si128(&xmm0);
+    xmm0 = _mm_maddubs_epi16(xmm0, *(__m128i*)kYuvConstants.kUVToB);
+    xmm1 = _mm_maddubs_epi16(xmm1, *(__m128i*)kYuvConstants.kUVToG);
+    xmm2 = _mm_maddubs_epi16(xmm2, *(__m128i*)kYuvConstants.kUVToR);
+    xmm0 = _mm_sub_epi16(*(__m128i*)kYuvConstants.kUVBiasB, xmm0);
+    xmm1 = _mm_sub_epi16(*(__m128i*)kYuvConstants.kUVBiasG, xmm1);
+    xmm2 = _mm_sub_epi16(*(__m128i*)kYuvConstants.kUVBiasR, xmm2);
     xmm3 = _mm_loadl_epi64((__m128i*)y_buf);
-    xmm3 = _mm_unpacklo_epi8(xmm3, xmm4);
-    xmm3 = _mm_subs_epi16(xmm3, *(__m128i*)kYSub16);
-    xmm3 = _mm_mullo_epi16(xmm3, *(__m128i*)kYToRgb);
+    xmm3 = _mm_unpacklo_epi8(xmm3, xmm3);
+    xmm3 = _mm_mulhi_epu16(xmm3, *(__m128i*)kYuvConstants.kYToRgb);
     xmm0 = _mm_adds_epi16(xmm0, xmm3);
     xmm1 = _mm_adds_epi16(xmm1, xmm3);
     xmm2 = _mm_adds_epi16(xmm2, xmm3);
@@ -112,60 +118,7 @@ void I422ToARGBRow_SSSE3(const uint8* y_buf,
     xmm2 = _mm_packus_epi16(xmm2, xmm2);
     xmm0 = _mm_unpacklo_epi8(xmm0, xmm1);
     xmm2 = _mm_unpacklo_epi8(xmm2, xmm5);
-    xmm1 = _mm_load_si128(&xmm0);
-    xmm0 = _mm_unpacklo_epi16(xmm0, xmm2);
-    xmm1 = _mm_unpackhi_epi16(xmm1, xmm2);
-
-    _mm_store_si128((__m128i *)dst_argb, xmm0);
-    _mm_store_si128((__m128i *)(dst_argb + 16), xmm1);
-
-    y_buf += 8;
-    u_buf += 4;
-    dst_argb += 32;
-    width -= 8;
-  }
-}
-
-// Unaligned destination version.
-void I422ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
-                                   const uint8* u_buf,
-                                   const uint8* v_buf,
-                                   uint8* dst_argb,
-                                   int width) {
-  __m128i xmm0, xmm1, xmm2, xmm3;
-  const __m128i xmm5 = _mm_set1_epi8(-1);
-  const __m128i xmm4 = _mm_setzero_si128();
-  const ptrdiff_t offset = (uint8*)v_buf - (uint8*)u_buf;
-
-  while (width > 0) {
-    xmm0 = _mm_cvtsi32_si128(*(uint32*)u_buf);
-    xmm1 = _mm_cvtsi32_si128(*(uint32*)(u_buf + offset));
-    xmm0 = _mm_unpacklo_epi8(xmm0, xmm1);
-    xmm0 = _mm_unpacklo_epi16(xmm0, xmm0);
-    xmm1 = _mm_load_si128(&xmm0);
-    xmm2 = _mm_load_si128(&xmm0);
-    xmm0 = _mm_maddubs_epi16(xmm0, *(__m128i*)kUVToB);
-    xmm1 = _mm_maddubs_epi16(xmm1, *(__m128i*)kUVToG);
-    xmm2 = _mm_maddubs_epi16(xmm2, *(__m128i*)kUVToR);
-    xmm0 = _mm_sub_epi16(xmm0, *(__m128i*)kUVBiasB);
-    xmm1 = _mm_sub_epi16(xmm1, *(__m128i*)kUVBiasG);
-    xmm2 = _mm_sub_epi16(xmm2, *(__m128i*)kUVBiasR);
-    xmm3 = _mm_loadl_epi64((__m128i*)y_buf);
-    xmm3 = _mm_unpacklo_epi8(xmm3, xmm4);
-    xmm3 = _mm_subs_epi16(xmm3, *(__m128i*)kYSub16);
-    xmm3 = _mm_mullo_epi16(xmm3, *(__m128i*)kYToRgb);
-    xmm0 = _mm_adds_epi16(xmm0, xmm3);
-    xmm1 = _mm_adds_epi16(xmm1, xmm3);
-    xmm2 = _mm_adds_epi16(xmm2, xmm3);
-    xmm0 = _mm_srai_epi16(xmm0, 6);
-    xmm1 = _mm_srai_epi16(xmm1, 6);
-    xmm2 = _mm_srai_epi16(xmm2, 6);
-    xmm0 = _mm_packus_epi16(xmm0, xmm0);
-    xmm1 = _mm_packus_epi16(xmm1, xmm1);
-    xmm2 = _mm_packus_epi16(xmm2, xmm2);
-    xmm0 = _mm_unpacklo_epi8(xmm0, xmm1);
-    xmm2 = _mm_unpacklo_epi8(xmm2, xmm5);
-    xmm1 = _mm_load_si128(&xmm0);
+    xmm1 = _mm_loadu_si128(&xmm0);
     xmm0 = _mm_unpacklo_epi16(xmm0, xmm2);
     xmm1 = _mm_unpackhi_epi16(xmm1, xmm2);
 
@@ -178,6 +131,7 @@ void I422ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
     width -= 8;
   }
 }
+
 // 32 bit
 #else  // defined(_M_X64)
 
@@ -209,15 +163,10 @@ static const vec8 kARGBToVJ = {
   -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0
 };
 
-// vpermd for vphaddw + vpackuswb vpermd.
-static const lvec32 kPermdARGBToY_AVX = {
-  0, 4, 1, 5, 2, 6, 3, 7
-};
-
 // vpshufb for vphaddw + vpackuswb packed to shorts.
 static const lvec8 kShufARGBToUV_AVX = {
   0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15,
-  0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15,
+  0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15
 };
 
 // Constants for BGRA.
@@ -263,6 +212,7 @@ static const uvec8 kAddY16 = {
   16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u
 };
 
+// 7 bit fixed point 0.5.
 static const vec16 kAddYJ64 = {
   64, 64, 64, 64, 64, 64, 64, 64
 };
@@ -316,36 +266,6 @@ void I400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix) {
     pcmpeqb    xmm5, xmm5            // generate mask 0xff000000
     pslld      xmm5, 24
 
-    align      4
-  convertloop:
-    movq       xmm0, qword ptr [eax]
-    lea        eax,  [eax + 8]
-    punpcklbw  xmm0, xmm0
-    movdqa     xmm1, xmm0
-    punpcklwd  xmm0, xmm0
-    punpckhwd  xmm1, xmm1
-    por        xmm0, xmm5
-    por        xmm1, xmm5
-    movdqa     [edx], xmm0
-    movdqa     [edx + 16], xmm1
-    lea        edx, [edx + 32]
-    sub        ecx, 8
-    jg         convertloop
-    ret
-  }
-}
-
-__declspec(naked) __declspec(align(16))
-void I400ToARGBRow_Unaligned_SSE2(const uint8* src_y, uint8* dst_argb,
-                                  int pix) {
-  __asm {
-    mov        eax, [esp + 4]        // src_y
-    mov        edx, [esp + 8]        // dst_argb
-    mov        ecx, [esp + 12]       // pix
-    pcmpeqb    xmm5, xmm5            // generate mask 0xff000000
-    pslld      xmm5, 24
-
-    align      4
   convertloop:
     movq       xmm0, qword ptr [eax]
     lea        eax,  [eax + 8]
@@ -374,7 +294,6 @@ void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix) {
     pslld     xmm5, 24
     movdqa    xmm4, kShuffleMaskRGB24ToARGB
 
-    align      4
  convertloop:
     movdqu    xmm0, [eax]
     movdqu    xmm1, [eax + 16]
@@ -386,18 +305,18 @@ void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix) {
     por       xmm2, xmm5
     palignr   xmm1, xmm0, 12   // xmm1 = { xmm3[0:7] xmm0[12:15]}
     pshufb    xmm0, xmm4
-    movdqa    [edx + 32], xmm2
+    movdqu    [edx + 32], xmm2
     por       xmm0, xmm5
     pshufb    xmm1, xmm4
-    movdqa    [edx], xmm0
+    movdqu    [edx], xmm0
     por       xmm1, xmm5
     palignr   xmm3, xmm3, 4    // xmm3 = { xmm3[4:15]}
     pshufb    xmm3, xmm4
-    movdqa    [edx + 16], xmm1
+    movdqu    [edx + 16], xmm1
     por       xmm3, xmm5
-    sub       ecx, 16
-    movdqa    [edx + 48], xmm3
+    movdqu    [edx + 48], xmm3
     lea       edx, [edx + 64]
+    sub       ecx, 16
     jg        convertloop
     ret
   }
@@ -414,7 +333,6 @@ void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb,
     pslld     xmm5, 24
     movdqa    xmm4, kShuffleMaskRAWToARGB
 
-    align      4
  convertloop:
     movdqu    xmm0, [eax]
     movdqu    xmm1, [eax + 16]
@@ -426,18 +344,18 @@ void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb,
     por       xmm2, xmm5
     palignr   xmm1, xmm0, 12   // xmm1 = { xmm3[0:7] xmm0[12:15]}
     pshufb    xmm0, xmm4
-    movdqa    [edx + 32], xmm2
+    movdqu    [edx + 32], xmm2
     por       xmm0, xmm5
     pshufb    xmm1, xmm4
-    movdqa    [edx], xmm0
+    movdqu    [edx], xmm0
     por       xmm1, xmm5
     palignr   xmm3, xmm3, 4    // xmm3 = { xmm3[4:15]}
     pshufb    xmm3, xmm4
-    movdqa    [edx + 16], xmm1
+    movdqu    [edx + 16], xmm1
     por       xmm3, xmm5
-    sub       ecx, 16
-    movdqa    [edx + 48], xmm3
+    movdqu    [edx + 48], xmm3
     lea       edx, [edx + 64]
+    sub       ecx, 16
     jg        convertloop
     ret
   }
@@ -474,7 +392,6 @@ void RGB565ToARGBRow_SSE2(const uint8* src_rgb565, uint8* dst_argb,
     sub       edx, eax
     sub       edx, eax
 
-    align      4
  convertloop:
     movdqu    xmm0, [eax]   // fetch 8 pixels of bgr565
     movdqa    xmm1, xmm0
@@ -491,8 +408,8 @@ void RGB565ToARGBRow_SSE2(const uint8* src_rgb565, uint8* dst_argb,
     movdqa    xmm2, xmm1
     punpcklbw xmm1, xmm0
     punpckhbw xmm2, xmm0
-    movdqa    [eax * 2 + edx], xmm1  // store 4 pixels of ARGB
-    movdqa    [eax * 2 + edx + 16], xmm2  // store next 4 pixels of ARGB
+    movdqu    [eax * 2 + edx], xmm1  // store 4 pixels of ARGB
+    movdqu    [eax * 2 + edx + 16], xmm2  // store next 4 pixels of ARGB
     lea       eax, [eax + 16]
     sub       ecx, 8
     jg        convertloop
@@ -524,7 +441,6 @@ void ARGB1555ToARGBRow_SSE2(const uint8* src_argb1555, uint8* dst_argb,
     sub       edx, eax
     sub       edx, eax
 
-    align      4
  convertloop:
     movdqu    xmm0, [eax]   // fetch 8 pixels of 1555
     movdqa    xmm1, xmm0
@@ -545,8 +461,8 @@ void ARGB1555ToARGBRow_SSE2(const uint8* src_argb1555, uint8* dst_argb,
     movdqa    xmm2, xmm1
     punpcklbw xmm1, xmm0
     punpckhbw xmm2, xmm0
-    movdqa    [eax * 2 + edx], xmm1  // store 4 pixels of ARGB
-    movdqa    [eax * 2 + edx + 16], xmm2  // store next 4 pixels of ARGB
+    movdqu    [eax * 2 + edx], xmm1  // store 4 pixels of ARGB
+    movdqu    [eax * 2 + edx + 16], xmm2  // store next 4 pixels of ARGB
     lea       eax, [eax + 16]
     sub       ecx, 8
     jg        convertloop
@@ -570,7 +486,6 @@ void ARGB4444ToARGBRow_SSE2(const uint8* src_argb4444, uint8* dst_argb,
     sub       edx, eax
     sub       edx, eax
 
-    align      4
  convertloop:
     movdqu    xmm0, [eax]   // fetch 8 pixels of bgra4444
     movdqa    xmm2, xmm0
@@ -585,8 +500,8 @@ void ARGB4444ToARGBRow_SSE2(const uint8* src_argb4444, uint8* dst_argb,
     movdqa    xmm1, xmm0
     punpcklbw xmm0, xmm2
     punpckhbw xmm1, xmm2
-    movdqa    [eax * 2 + edx], xmm0  // store 4 pixels of ARGB
-    movdqa    [eax * 2 + edx + 16], xmm1  // store next 4 pixels of ARGB
+    movdqu    [eax * 2 + edx], xmm0  // store 4 pixels of ARGB
+    movdqu    [eax * 2 + edx + 16], xmm1  // store next 4 pixels of ARGB
     lea       eax, [eax + 16]
     sub       ecx, 8
     jg        convertloop
@@ -602,7 +517,6 @@ void ARGBToRGB24Row_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix) {
     mov       ecx, [esp + 12]  // pix
     movdqa    xmm6, kShuffleMaskARGBToRGB24
 
-    align      4
  convertloop:
     movdqu    xmm0, [eax]   // fetch 16 pixels of argb
     movdqu    xmm1, [eax + 16]
@@ -641,7 +555,6 @@ void ARGBToRAWRow_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix) {
     mov       ecx, [esp + 12]  // pix
     movdqa    xmm6, kShuffleMaskARGBToRAW
 
-    align      4
  convertloop:
     movdqu    xmm0, [eax]   // fetch 16 pixels of argb
     movdqu    xmm1, [eax + 16]
@@ -686,9 +599,8 @@ void ARGBToRGB565Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix) {
     pcmpeqb   xmm5, xmm5       // generate mask 0xfffff800
     pslld     xmm5, 11
 
-    align      4
  convertloop:
-    movdqa    xmm0, [eax]   // fetch 4 pixels of argb
+    movdqu    xmm0, [eax]   // fetch 4 pixels of argb
     movdqa    xmm1, xmm0    // B
     movdqa    xmm2, xmm0    // G
     pslld     xmm0, 8       // R
@@ -726,9 +638,8 @@ void ARGBToARGB1555Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix) {
     pcmpeqb   xmm7, xmm7       // generate mask 0xffff8000
     pslld     xmm7, 15
 
-    align      4
  convertloop:
-    movdqa    xmm0, [eax]   // fetch 4 pixels of argb
+    movdqu    xmm0, [eax]   // fetch 4 pixels of argb
     movdqa    xmm1, xmm0    // B
     movdqa    xmm2, xmm0    // G
     movdqa    xmm3, xmm0    // R
@@ -764,14 +675,13 @@ void ARGBToARGB4444Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix) {
     movdqa    xmm3, xmm4       // generate mask 0x00f000f0
     psrlw     xmm3, 8
 
-    align      4
  convertloop:
-    movdqa    xmm0, [eax]   // fetch 4 pixels of argb
+    movdqu    xmm0, [eax]   // fetch 4 pixels of argb
     movdqa    xmm1, xmm0
     pand      xmm0, xmm3    // low nibble
     pand      xmm1, xmm4    // high nibble
-    psrl      xmm0, 4
-    psrl      xmm1, 8
+    psrld     xmm0, 4
+    psrld     xmm1, 8
     por       xmm0, xmm1
     packuswb  xmm0, xmm0
     lea       eax, [eax + 16]
@@ -783,6 +693,116 @@ void ARGBToARGB4444Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix) {
   }
 }
 
+#ifdef HAS_ARGBTORGB565ROW_AVX2
+__declspec(naked) __declspec(align(16))
+void ARGBToRGB565Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov        eax, [esp + 4]      // src_argb
+    mov        edx, [esp + 8]      // dst_rgb
+    mov        ecx, [esp + 12]     // pix
+    vpcmpeqb   ymm3, ymm3, ymm3    // generate mask 0x0000001f
+    vpsrld     ymm3, ymm3, 27
+    vpcmpeqb   ymm4, ymm4, ymm4    // generate mask 0x000007e0
+    vpsrld     ymm4, ymm4, 26
+    vpslld     ymm4, ymm4, 5
+    vpcmpeqb   ymm5, ymm5, ymm5    // generate mask 0xfffff800
+    vpslld     ymm5, ymm5, 11
+
+ convertloop:
+    vmovdqu    ymm0, [eax]         // fetch 8 pixels of argb
+    vpsrld     ymm2, ymm0, 5       // G
+    vpsrld     ymm1, ymm0, 3       // B
+    vpslld     ymm0, ymm0, 8       // R
+    vpand      ymm2, ymm2, ymm4    // G
+    vpand      ymm1, ymm1, ymm3    // B
+    vpsrad     ymm0, ymm0, 16      // R
+    vpand      ymm0, ymm0, ymm5    // R
+    vpor       ymm1, ymm1, ymm2    // BG
+    vpor       ymm0, ymm0, ymm1    // BGR
+    vpackssdw  ymm0, ymm0, ymm0
+    vpermq     ymm0, ymm0, 0xd8
+    lea        eax, [eax + 32]
+    vmovdqu    [edx], xmm0         // store 8 pixels of RGB565
+    lea        edx, [edx + 16]
+    sub        ecx, 8
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBTORGB565ROW_AVX2
+
+#ifdef HAS_ARGBTOARGB1555ROW_AVX2
+__declspec(naked) __declspec(align(16))
+void ARGBToARGB1555Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov        eax, [esp + 4]      // src_argb
+    mov        edx, [esp + 8]      // dst_rgb
+    mov        ecx, [esp + 12]     // pix
+    vpcmpeqb   ymm4, ymm4, ymm4
+    vpsrld     ymm4, ymm4, 27      // generate mask 0x0000001f
+    vpslld     ymm5, ymm4, 5       // generate mask 0x000003e0
+    vpslld     ymm6, ymm4, 10      // generate mask 0x00007c00
+    vpcmpeqb   ymm7, ymm7, ymm7    // generate mask 0xffff8000
+    vpslld     ymm7, ymm7, 15
+
+ convertloop:
+    vmovdqu    ymm0, [eax]         // fetch 8 pixels of argb
+    vpsrld     ymm3, ymm0, 9       // R
+    vpsrld     ymm2, ymm0, 6       // G
+    vpsrld     ymm1, ymm0, 3       // B
+    vpsrad     ymm0, ymm0, 16      // A
+    vpand      ymm3, ymm3, ymm6    // R
+    vpand      ymm2, ymm2, ymm5    // G
+    vpand      ymm1, ymm1, ymm4    // B
+    vpand      ymm0, ymm0, ymm7    // A
+    vpor       ymm0, ymm0, ymm1    // BA
+    vpor       ymm2, ymm2, ymm3    // GR
+    vpor       ymm0, ymm0, ymm2    // BGRA
+    vpackssdw  ymm0, ymm0, ymm0
+    vpermq     ymm0, ymm0, 0xd8
+    lea        eax, [eax + 32]
+    vmovdqu    [edx], xmm0         // store 8 pixels of ARGB1555
+    lea        edx, [edx + 16]
+    sub        ecx, 8
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBTOARGB1555ROW_AVX2
+
+#ifdef HAS_ARGBTOARGB4444ROW_AVX2
+__declspec(naked) __declspec(align(16))
+void ARGBToARGB4444Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb
+    mov        edx, [esp + 8]   // dst_rgb
+    mov        ecx, [esp + 12]  // pix
+    vpcmpeqb   ymm4, ymm4, ymm4   // generate mask 0xf000f000
+    vpsllw     ymm4, ymm4, 12
+    vpsrlw     ymm3, ymm4, 8      // generate mask 0x00f000f0
+
+ convertloop:
+    vmovdqu    ymm0, [eax]         // fetch 8 pixels of argb
+    vpand      ymm1, ymm0, ymm4    // high nibble
+    vpand      ymm0, ymm0, ymm3    // low nibble
+    vpsrld     ymm1, ymm1, 8
+    vpsrld     ymm0, ymm0, 4
+    vpor       ymm0, ymm0, ymm1
+    vpackuswb  ymm0, ymm0, ymm0
+    vpermq     ymm0, ymm0, 0xd8
+    lea        eax, [eax + 32]
+    vmovdqu    [edx], xmm0         // store 8 pixels of ARGB4444
+    lea        edx, [edx + 16]
+    sub        ecx, 8
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBTOARGB4444ROW_AVX2
+
 // Convert 16 ARGB pixels (64 bytes) to 16 Y values.
 __declspec(naked) __declspec(align(16))
 void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
@@ -790,15 +810,14 @@ void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
     mov        eax, [esp + 4]   /* src_argb */
     mov        edx, [esp + 8]   /* dst_y */
     mov        ecx, [esp + 12]  /* pix */
-    movdqa     xmm5, kAddY16
     movdqa     xmm4, kARGBToY
+    movdqa     xmm5, kAddY16
 
-    align      4
  convertloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    movdqa     xmm2, [eax + 32]
-    movdqa     xmm3, [eax + 48]
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
     pmaddubsw  xmm0, xmm4
     pmaddubsw  xmm1, xmm4
     pmaddubsw  xmm2, xmm4
@@ -810,15 +829,16 @@ void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
     psrlw      xmm2, 7
     packuswb   xmm0, xmm2
     paddb      xmm0, xmm5
-    sub        ecx, 16
-    movdqa     [edx], xmm0
+    movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 16
     jg         convertloop
     ret
   }
 }
 
-// Convert 16 ARGB pixels (64 bytes) to 16 Y values.
+// Convert 16 ARGB pixels (64 bytes) to 16 YJ values.
+// Same as ARGBToYRow but different coefficients, no add 16, but do rounding.
 __declspec(naked) __declspec(align(16))
 void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
   __asm {
@@ -828,12 +848,11 @@ void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
     movdqa     xmm4, kARGBToYJ
     movdqa     xmm5, kAddYJ64
 
-    align      4
  convertloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    movdqa     xmm2, [eax + 32]
-    movdqa     xmm3, [eax + 48]
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
     pmaddubsw  xmm0, xmm4
     pmaddubsw  xmm1, xmm4
     pmaddubsw  xmm2, xmm4
@@ -846,15 +865,20 @@ void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
     psrlw      xmm0, 7
     psrlw      xmm2, 7
     packuswb   xmm0, xmm2
-    sub        ecx, 16
-    movdqa     [edx], xmm0
+    movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 16
     jg         convertloop
     ret
   }
 }
 
 #ifdef HAS_ARGBTOYROW_AVX2
+// vpermd for vphaddw + vpackuswb vpermd.
+static const lvec32 kPermdARGBToY_AVX = {
+  0, 4, 1, 5, 2, 6, 3, 7
+};
+
 // Convert 32 ARGB pixels (128 bytes) to 32 Y values.
 __declspec(naked) __declspec(align(32))
 void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) {
@@ -864,9 +888,8 @@ void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) {
     mov        ecx, [esp + 12]  /* pix */
     vbroadcastf128 ymm4, kARGBToY
     vbroadcastf128 ymm5, kAddY16
-    vmovdqa    ymm6, kPermdARGBToY_AVX
+    vmovdqu    ymm6, kPermdARGBToY_AVX
 
-    align      4
  convertloop:
     vmovdqu    ymm0, [eax]
     vmovdqu    ymm1, [eax + 32]
@@ -883,10 +906,10 @@ void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) {
     vpsrlw     ymm2, ymm2, 7
     vpackuswb  ymm0, ymm0, ymm2  // mutates.
     vpermd     ymm0, ymm6, ymm0  // For vphaddw + vpackuswb mutation.
-    vpaddb     ymm0, ymm0, ymm5
-    sub        ecx, 32
+    vpaddb     ymm0, ymm0, ymm5  // add 16 for Y
     vmovdqu    [edx], ymm0
     lea        edx, [edx + 32]
+    sub        ecx, 32
     jg         convertloop
     vzeroupper
     ret
@@ -904,9 +927,8 @@ void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) {
     mov        ecx, [esp + 12]  /* pix */
     vbroadcastf128 ymm4, kARGBToYJ
     vbroadcastf128 ymm5, kAddYJ64
-    vmovdqa    ymm6, kPermdARGBToY_AVX
+    vmovdqu    ymm6, kPermdARGBToY_AVX
 
-    align      4
  convertloop:
     vmovdqu    ymm0, [eax]
     vmovdqu    ymm1, [eax + 32]
@@ -925,9 +947,9 @@ void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) {
     vpsrlw     ymm2, ymm2, 7
     vpackuswb  ymm0, ymm0, ymm2  // mutates.
     vpermd     ymm0, ymm6, ymm0  // For vphaddw + vpackuswb mutation.
-    sub        ecx, 32
     vmovdqu    [edx], ymm0
     lea        edx, [edx + 32]
+    sub        ecx, 32
     jg         convertloop
 
     vzeroupper
@@ -937,118 +959,14 @@ void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) {
 #endif  //  HAS_ARGBTOYJROW_AVX2
 
 __declspec(naked) __declspec(align(16))
-void ARGBToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
-  __asm {
-    mov        eax, [esp + 4]   /* src_argb */
-    mov        edx, [esp + 8]   /* dst_y */
-    mov        ecx, [esp + 12]  /* pix */
-    movdqa     xmm5, kAddY16
-    movdqa     xmm4, kARGBToY
-
-    align      4
- convertloop:
-    movdqu     xmm0, [eax]
-    movdqu     xmm1, [eax + 16]
-    movdqu     xmm2, [eax + 32]
-    movdqu     xmm3, [eax + 48]
-    pmaddubsw  xmm0, xmm4
-    pmaddubsw  xmm1, xmm4
-    pmaddubsw  xmm2, xmm4
-    pmaddubsw  xmm3, xmm4
-    lea        eax, [eax + 64]
-    phaddw     xmm0, xmm1
-    phaddw     xmm2, xmm3
-    psrlw      xmm0, 7
-    psrlw      xmm2, 7
-    packuswb   xmm0, xmm2
-    paddb      xmm0, xmm5
-    sub        ecx, 16
-    movdqu     [edx], xmm0
-    lea        edx, [edx + 16]
-    jg         convertloop
-    ret
-  }
-}
-
-__declspec(naked) __declspec(align(16))
-void ARGBToYJRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
-  __asm {
-    mov        eax, [esp + 4]   /* src_argb */
-    mov        edx, [esp + 8]   /* dst_y */
-    mov        ecx, [esp + 12]  /* pix */
-    movdqa     xmm4, kARGBToYJ
-    movdqa     xmm5, kAddYJ64
-
-    align      4
- convertloop:
-    movdqu     xmm0, [eax]
-    movdqu     xmm1, [eax + 16]
-    movdqu     xmm2, [eax + 32]
-    movdqu     xmm3, [eax + 48]
-    pmaddubsw  xmm0, xmm4
-    pmaddubsw  xmm1, xmm4
-    pmaddubsw  xmm2, xmm4
-    pmaddubsw  xmm3, xmm4
-    lea        eax, [eax + 64]
-    phaddw     xmm0, xmm1
-    phaddw     xmm2, xmm3
-    paddw      xmm0, xmm5
-    paddw      xmm2, xmm5
-    psrlw      xmm0, 7
-    psrlw      xmm2, 7
-    packuswb   xmm0, xmm2
-    sub        ecx, 16
-    movdqu     [edx], xmm0
-    lea        edx, [edx + 16]
-    jg         convertloop
-    ret
-  }
-}
-
-__declspec(naked) __declspec(align(16))
 void BGRAToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
   __asm {
     mov        eax, [esp + 4]   /* src_argb */
     mov        edx, [esp + 8]   /* dst_y */
     mov        ecx, [esp + 12]  /* pix */
-    movdqa     xmm5, kAddY16
     movdqa     xmm4, kBGRAToY
-
-    align      4
- convertloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    movdqa     xmm2, [eax + 32]
-    movdqa     xmm3, [eax + 48]
-    pmaddubsw  xmm0, xmm4
-    pmaddubsw  xmm1, xmm4
-    pmaddubsw  xmm2, xmm4
-    pmaddubsw  xmm3, xmm4
-    lea        eax, [eax + 64]
-    phaddw     xmm0, xmm1
-    phaddw     xmm2, xmm3
-    psrlw      xmm0, 7
-    psrlw      xmm2, 7
-    packuswb   xmm0, xmm2
-    paddb      xmm0, xmm5
-    sub        ecx, 16
-    movdqa     [edx], xmm0
-    lea        edx, [edx + 16]
-    jg         convertloop
-    ret
-  }
-}
-
-__declspec(naked) __declspec(align(16))
-void BGRAToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
-  __asm {
-    mov        eax, [esp + 4]   /* src_argb */
-    mov        edx, [esp + 8]   /* dst_y */
-    mov        ecx, [esp + 12]  /* pix */
     movdqa     xmm5, kAddY16
-    movdqa     xmm4, kBGRAToY
 
-    align      4
  convertloop:
     movdqu     xmm0, [eax]
     movdqu     xmm1, [eax + 16]
@@ -1065,9 +983,9 @@ void BGRAToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
     psrlw      xmm2, 7
     packuswb   xmm0, xmm2
     paddb      xmm0, xmm5
-    sub        ecx, 16
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 16
     jg         convertloop
     ret
   }
@@ -1079,44 +997,9 @@ void ABGRToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
     mov        eax, [esp + 4]   /* src_argb */
     mov        edx, [esp + 8]   /* dst_y */
     mov        ecx, [esp + 12]  /* pix */
-    movdqa     xmm5, kAddY16
     movdqa     xmm4, kABGRToY
-
-    align      4
- convertloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    movdqa     xmm2, [eax + 32]
-    movdqa     xmm3, [eax + 48]
-    pmaddubsw  xmm0, xmm4
-    pmaddubsw  xmm1, xmm4
-    pmaddubsw  xmm2, xmm4
-    pmaddubsw  xmm3, xmm4
-    lea        eax, [eax + 64]
-    phaddw     xmm0, xmm1
-    phaddw     xmm2, xmm3
-    psrlw      xmm0, 7
-    psrlw      xmm2, 7
-    packuswb   xmm0, xmm2
-    paddb      xmm0, xmm5
-    sub        ecx, 16
-    movdqa     [edx], xmm0
-    lea        edx, [edx + 16]
-    jg         convertloop
-    ret
-  }
-}
-
-__declspec(naked) __declspec(align(16))
-void ABGRToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
-  __asm {
-    mov        eax, [esp + 4]   /* src_argb */
-    mov        edx, [esp + 8]   /* dst_y */
-    mov        ecx, [esp + 12]  /* pix */
     movdqa     xmm5, kAddY16
-    movdqa     xmm4, kABGRToY
 
-    align      4
  convertloop:
     movdqu     xmm0, [eax]
     movdqu     xmm1, [eax + 16]
@@ -1133,9 +1016,9 @@ void ABGRToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
     psrlw      xmm2, 7
     packuswb   xmm0, xmm2
     paddb      xmm0, xmm5
-    sub        ecx, 16
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 16
     jg         convertloop
     ret
   }
@@ -1147,44 +1030,9 @@ void RGBAToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
     mov        eax, [esp + 4]   /* src_argb */
     mov        edx, [esp + 8]   /* dst_y */
     mov        ecx, [esp + 12]  /* pix */
-    movdqa     xmm5, kAddY16
     movdqa     xmm4, kRGBAToY
-
-    align      4
- convertloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    movdqa     xmm2, [eax + 32]
-    movdqa     xmm3, [eax + 48]
-    pmaddubsw  xmm0, xmm4
-    pmaddubsw  xmm1, xmm4
-    pmaddubsw  xmm2, xmm4
-    pmaddubsw  xmm3, xmm4
-    lea        eax, [eax + 64]
-    phaddw     xmm0, xmm1
-    phaddw     xmm2, xmm3
-    psrlw      xmm0, 7
-    psrlw      xmm2, 7
-    packuswb   xmm0, xmm2
-    paddb      xmm0, xmm5
-    sub        ecx, 16
-    movdqa     [edx], xmm0
-    lea        edx, [edx + 16]
-    jg         convertloop
-    ret
-  }
-}
-
-__declspec(naked) __declspec(align(16))
-void RGBAToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
-  __asm {
-    mov        eax, [esp + 4]   /* src_argb */
-    mov        edx, [esp + 8]   /* dst_y */
-    mov        ecx, [esp + 12]  /* pix */
     movdqa     xmm5, kAddY16
-    movdqa     xmm4, kRGBAToY
 
-    align      4
  convertloop:
     movdqu     xmm0, [eax]
     movdqu     xmm1, [eax + 16]
@@ -1201,9 +1049,9 @@ void RGBAToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
     psrlw      xmm2, 7
     packuswb   xmm0, xmm2
     paddb      xmm0, xmm5
-    sub        ecx, 16
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 16
     jg         convertloop
     ret
   }
@@ -1220,22 +1068,26 @@ void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
     mov        edx, [esp + 8 + 12]  // dst_u
     mov        edi, [esp + 8 + 16]  // dst_v
     mov        ecx, [esp + 8 + 20]  // pix
-    movdqa     xmm7, kARGBToU
-    movdqa     xmm6, kARGBToV
     movdqa     xmm5, kAddUV128
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm7, kARGBToU
     sub        edi, edx             // stride from u to v
 
-    align      4
  convertloop:
     /* step 1 - subsample 16x2 argb pixels to 8x1 */
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    movdqa     xmm2, [eax + 32]
-    movdqa     xmm3, [eax + 48]
-    pavgb      xmm0, [eax + esi]
-    pavgb      xmm1, [eax + esi + 16]
-    pavgb      xmm2, [eax + esi + 32]
-    pavgb      xmm3, [eax + esi + 48]
+    movdqu     xmm0, [eax]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+
     lea        eax,  [eax + 64]
     movdqa     xmm4, xmm0
     shufps     xmm0, xmm1, 0x88
@@ -1263,10 +1115,10 @@ void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
     paddb      xmm0, xmm5            // -> unsigned
 
     // step 3 - store 8 U and 8 V values
-    sub        ecx, 16
     movlps     qword ptr [edx], xmm0 // U
     movhps     qword ptr [edx + edi], xmm0 // V
     lea        edx, [edx + 8]
+    sub        ecx, 16
     jg         convertloop
 
     pop        edi
@@ -1286,22 +1138,26 @@ void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
     mov        edx, [esp + 8 + 12]  // dst_u
     mov        edi, [esp + 8 + 16]  // dst_v
     mov        ecx, [esp + 8 + 20]  // pix
-    movdqa     xmm7, kARGBToUJ
-    movdqa     xmm6, kARGBToVJ
     movdqa     xmm5, kAddUVJ128
+    movdqa     xmm6, kARGBToVJ
+    movdqa     xmm7, kARGBToUJ
     sub        edi, edx             // stride from u to v
 
-    align      4
  convertloop:
     /* step 1 - subsample 16x2 argb pixels to 8x1 */
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    movdqa     xmm2, [eax + 32]
-    movdqa     xmm3, [eax + 48]
-    pavgb      xmm0, [eax + esi]
-    pavgb      xmm1, [eax + esi + 16]
-    pavgb      xmm2, [eax + esi + 32]
-    pavgb      xmm3, [eax + esi + 48]
+    movdqu     xmm0, [eax]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+
     lea        eax,  [eax + 64]
     movdqa     xmm4, xmm0
     shufps     xmm0, xmm1, 0x88
@@ -1330,10 +1186,10 @@ void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
     packsswb   xmm0, xmm1
 
     // step 3 - store 8 U and 8 V values
-    sub        ecx, 16
     movlps     qword ptr [edx], xmm0 // U
     movhps     qword ptr [edx + edi], xmm0 // V
     lea        edx, [edx + 8]
+    sub        ecx, 16
     jg         convertloop
 
     pop        edi
@@ -1359,7 +1215,6 @@ void ARGBToUVRow_AVX2(const uint8* src_argb0, int src_stride_argb,
     vbroadcastf128 ymm7, kARGBToU
     sub        edi, edx             // stride from u to v
 
-    align      4
  convertloop:
     /* step 1 - subsample 32x2 argb pixels to 16x1 */
     vmovdqu    ymm0, [eax]
@@ -1395,10 +1250,10 @@ void ARGBToUVRow_AVX2(const uint8* src_argb0, int src_stride_argb,
     vpaddb     ymm0, ymm0, ymm5  // -> unsigned
 
     // step 3 - store 16 U and 16 V values
-    sub         ecx, 32
     vextractf128 [edx], ymm0, 0 // U
     vextractf128 [edx + edi], ymm0, 1 // V
     lea        edx, [edx + 16]
+    sub        ecx, 32
     jg         convertloop
 
     pop        edi
@@ -1410,147 +1265,6 @@ void ARGBToUVRow_AVX2(const uint8* src_argb0, int src_stride_argb,
 #endif  // HAS_ARGBTOUVROW_AVX2
 
 __declspec(naked) __declspec(align(16))
-void ARGBToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                                 uint8* dst_u, uint8* dst_v, int width) {
-  __asm {
-    push       esi
-    push       edi
-    mov        eax, [esp + 8 + 4]   // src_argb
-    mov        esi, [esp + 8 + 8]   // src_stride_argb
-    mov        edx, [esp + 8 + 12]  // dst_u
-    mov        edi, [esp + 8 + 16]  // dst_v
-    mov        ecx, [esp + 8 + 20]  // pix
-    movdqa     xmm7, kARGBToU
-    movdqa     xmm6, kARGBToV
-    movdqa     xmm5, kAddUV128
-    sub        edi, edx             // stride from u to v
-
-    align      4
- convertloop:
-    /* step 1 - subsample 16x2 argb pixels to 8x1 */
-    movdqu     xmm0, [eax]
-    movdqu     xmm1, [eax + 16]
-    movdqu     xmm2, [eax + 32]
-    movdqu     xmm3, [eax + 48]
-    movdqu     xmm4, [eax + esi]
-    pavgb      xmm0, xmm4
-    movdqu     xmm4, [eax + esi + 16]
-    pavgb      xmm1, xmm4
-    movdqu     xmm4, [eax + esi + 32]
-    pavgb      xmm2, xmm4
-    movdqu     xmm4, [eax + esi + 48]
-    pavgb      xmm3, xmm4
-    lea        eax,  [eax + 64]
-    movdqa     xmm4, xmm0
-    shufps     xmm0, xmm1, 0x88
-    shufps     xmm4, xmm1, 0xdd
-    pavgb      xmm0, xmm4
-    movdqa     xmm4, xmm2
-    shufps     xmm2, xmm3, 0x88
-    shufps     xmm4, xmm3, 0xdd
-    pavgb      xmm2, xmm4
-
-    // step 2 - convert to U and V
-    // from here down is very similar to Y code except
-    // instead of 16 different pixels, its 8 pixels of U and 8 of V
-    movdqa     xmm1, xmm0
-    movdqa     xmm3, xmm2
-    pmaddubsw  xmm0, xmm7  // U
-    pmaddubsw  xmm2, xmm7
-    pmaddubsw  xmm1, xmm6  // V
-    pmaddubsw  xmm3, xmm6
-    phaddw     xmm0, xmm2
-    phaddw     xmm1, xmm3
-    psraw      xmm0, 8
-    psraw      xmm1, 8
-    packsswb   xmm0, xmm1
-    paddb      xmm0, xmm5            // -> unsigned
-
-    // step 3 - store 8 U and 8 V values
-    sub        ecx, 16
-    movlps     qword ptr [edx], xmm0 // U
-    movhps     qword ptr [edx + edi], xmm0 // V
-    lea        edx, [edx + 8]
-    jg         convertloop
-
-    pop        edi
-    pop        esi
-    ret
-  }
-}
-
-__declspec(naked) __declspec(align(16))
-void ARGBToUVJRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                                 uint8* dst_u, uint8* dst_v, int width) {
-  __asm {
-    push       esi
-    push       edi
-    mov        eax, [esp + 8 + 4]   // src_argb
-    mov        esi, [esp + 8 + 8]   // src_stride_argb
-    mov        edx, [esp + 8 + 12]  // dst_u
-    mov        edi, [esp + 8 + 16]  // dst_v
-    mov        ecx, [esp + 8 + 20]  // pix
-    movdqa     xmm7, kARGBToUJ
-    movdqa     xmm6, kARGBToVJ
-    movdqa     xmm5, kAddUVJ128
-    sub        edi, edx             // stride from u to v
-
-    align      4
- convertloop:
-    /* step 1 - subsample 16x2 argb pixels to 8x1 */
-    movdqu     xmm0, [eax]
-    movdqu     xmm1, [eax + 16]
-    movdqu     xmm2, [eax + 32]
-    movdqu     xmm3, [eax + 48]
-    movdqu     xmm4, [eax + esi]
-    pavgb      xmm0, xmm4
-    movdqu     xmm4, [eax + esi + 16]
-    pavgb      xmm1, xmm4
-    movdqu     xmm4, [eax + esi + 32]
-    pavgb      xmm2, xmm4
-    movdqu     xmm4, [eax + esi + 48]
-    pavgb      xmm3, xmm4
-    lea        eax,  [eax + 64]
-    movdqa     xmm4, xmm0
-    shufps     xmm0, xmm1, 0x88
-    shufps     xmm4, xmm1, 0xdd
-    pavgb      xmm0, xmm4
-    movdqa     xmm4, xmm2
-    shufps     xmm2, xmm3, 0x88
-    shufps     xmm4, xmm3, 0xdd
-    pavgb      xmm2, xmm4
-
-    // step 2 - convert to U and V
-    // from here down is very similar to Y code except
-    // instead of 16 different pixels, its 8 pixels of U and 8 of V
-    movdqa     xmm1, xmm0
-    movdqa     xmm3, xmm2
-    pmaddubsw  xmm0, xmm7  // U
-    pmaddubsw  xmm2, xmm7
-    pmaddubsw  xmm1, xmm6  // V
-    pmaddubsw  xmm3, xmm6
-    phaddw     xmm0, xmm2
-    phaddw     xmm1, xmm3
-    paddw      xmm0, xmm5            // +.5 rounding -> unsigned
-    paddw      xmm1, xmm5
-    psraw      xmm0, 8
-    psraw      xmm1, 8
-    packsswb   xmm0, xmm1
-
-    // step 3 - store 8 U and 8 V values
-    sub        ecx, 16
-    movlps     qword ptr [edx], xmm0 // U
-    movhps     qword ptr [edx + edi], xmm0 // V
-    lea        edx, [edx + 8]
-    jg         convertloop
-
-    pop        edi
-    pop        esi
-    ret
-  }
-}
-
-__declspec(naked) __declspec(align(16))
 void ARGBToUV444Row_SSSE3(const uint8* src_argb0,
                           uint8* dst_u, uint8* dst_v, int width) {
   __asm {
@@ -1559,70 +1273,11 @@ void ARGBToUV444Row_SSSE3(const uint8* src_argb0,
     mov        edx, [esp + 4 + 8]   // dst_u
     mov        edi, [esp + 4 + 12]  // dst_v
     mov        ecx, [esp + 4 + 16]  // pix
-    movdqa     xmm7, kARGBToU
-    movdqa     xmm6, kARGBToV
     movdqa     xmm5, kAddUV128
-    sub        edi, edx             // stride from u to v
-
-    align      4
- convertloop:
-    /* convert to U and V */
-    movdqa     xmm0, [eax]          // U
-    movdqa     xmm1, [eax + 16]
-    movdqa     xmm2, [eax + 32]
-    movdqa     xmm3, [eax + 48]
-    pmaddubsw  xmm0, xmm7
-    pmaddubsw  xmm1, xmm7
-    pmaddubsw  xmm2, xmm7
-    pmaddubsw  xmm3, xmm7
-    phaddw     xmm0, xmm1
-    phaddw     xmm2, xmm3
-    psraw      xmm0, 8
-    psraw      xmm2, 8
-    packsswb   xmm0, xmm2
-    paddb      xmm0, xmm5
-    sub        ecx,  16
-    movdqa     [edx], xmm0
-
-    movdqa     xmm0, [eax]          // V
-    movdqa     xmm1, [eax + 16]
-    movdqa     xmm2, [eax + 32]
-    movdqa     xmm3, [eax + 48]
-    pmaddubsw  xmm0, xmm6
-    pmaddubsw  xmm1, xmm6
-    pmaddubsw  xmm2, xmm6
-    pmaddubsw  xmm3, xmm6
-    phaddw     xmm0, xmm1
-    phaddw     xmm2, xmm3
-    psraw      xmm0, 8
-    psraw      xmm2, 8
-    packsswb   xmm0, xmm2
-    paddb      xmm0, xmm5
-    lea        eax,  [eax + 64]
-    movdqa     [edx + edi], xmm0
-    lea        edx,  [edx + 16]
-    jg         convertloop
-
-    pop        edi
-    ret
-  }
-}
-
-__declspec(naked) __declspec(align(16))
-void ARGBToUV444Row_Unaligned_SSSE3(const uint8* src_argb0,
-                                    uint8* dst_u, uint8* dst_v, int width) {
-  __asm {
-    push       edi
-    mov        eax, [esp + 4 + 4]   // src_argb
-    mov        edx, [esp + 4 + 8]   // dst_u
-    mov        edi, [esp + 4 + 12]  // dst_v
-    mov        ecx, [esp + 4 + 16]  // pix
-    movdqa     xmm7, kARGBToU
     movdqa     xmm6, kARGBToV
-    movdqa     xmm5, kAddUV128
+    movdqa     xmm7, kARGBToU
     sub        edi, edx             // stride from u to v
 
-    align      4
  convertloop:
     /* convert to U and V */
     movdqu     xmm0, [eax]          // U
@@ -1639,7 +1294,6 @@ void ARGBToUV444Row_Unaligned_SSSE3(const uint8* src_argb0,
     psraw      xmm2, 8
     packsswb   xmm0, xmm2
     paddb      xmm0, xmm5
-    sub        ecx,  16
     movdqu     [edx], xmm0
 
     movdqu     xmm0, [eax]          // V
@@ -1659,6 +1313,7 @@ void ARGBToUV444Row_Unaligned_SSSE3(const uint8* src_argb0,
     lea        eax,  [eax + 64]
     movdqu     [edx + edi], xmm0
     lea        edx,  [edx + 16]
+    sub        ecx,  16
     jg         convertloop
 
     pop        edi
@@ -1675,18 +1330,17 @@ void ARGBToUV422Row_SSSE3(const uint8* src_argb0,
     mov        edx, [esp + 4 + 8]   // dst_u
     mov        edi, [esp + 4 + 12]  // dst_v
     mov        ecx, [esp + 4 + 16]  // pix
-    movdqa     xmm7, kARGBToU
-    movdqa     xmm6, kARGBToV
     movdqa     xmm5, kAddUV128
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm7, kARGBToU
     sub        edi, edx             // stride from u to v
 
-    align      4
  convertloop:
     /* step 1 - subsample 16x2 argb pixels to 8x1 */
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    movdqa     xmm2, [eax + 32]
-    movdqa     xmm3, [eax + 48]
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
     lea        eax,  [eax + 64]
     movdqa     xmm4, xmm0
     shufps     xmm0, xmm1, 0x88
@@ -1714,10 +1368,10 @@ void ARGBToUV422Row_SSSE3(const uint8* src_argb0,
     paddb      xmm0, xmm5            // -> unsigned
 
     // step 3 - store 8 U and 8 V values
-    sub        ecx, 16
     movlps     qword ptr [edx], xmm0 // U
     movhps     qword ptr [edx + edi], xmm0 // V
     lea        edx, [edx + 8]
+    sub        ecx, 16
     jg         convertloop
 
     pop        edi
@@ -1726,26 +1380,36 @@ void ARGBToUV422Row_SSSE3(const uint8* src_argb0,
 }
 
 __declspec(naked) __declspec(align(16))
-void ARGBToUV422Row_Unaligned_SSSE3(const uint8* src_argb0,
-                                    uint8* dst_u, uint8* dst_v, int width) {
+void BGRAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
   __asm {
+    push       esi
     push       edi
-    mov        eax, [esp + 4 + 4]   // src_argb
-    mov        edx, [esp + 4 + 8]   // dst_u
-    mov        edi, [esp + 4 + 12]  // dst_v
-    mov        ecx, [esp + 4 + 16]  // pix
-    movdqa     xmm7, kARGBToU
-    movdqa     xmm6, kARGBToV
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
     movdqa     xmm5, kAddUV128
+    movdqa     xmm6, kBGRAToV
+    movdqa     xmm7, kBGRAToU
     sub        edi, edx             // stride from u to v
 
-    align      4
  convertloop:
     /* step 1 - subsample 16x2 argb pixels to 8x1 */
     movdqu     xmm0, [eax]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
     movdqu     xmm1, [eax + 16]
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
     movdqu     xmm2, [eax + 32]
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
     movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+
     lea        eax,  [eax + 64]
     movdqa     xmm4, xmm0
     shufps     xmm0, xmm1, 0x88
@@ -1773,19 +1437,20 @@ void ARGBToUV422Row_Unaligned_SSSE3(const uint8* src_argb0,
     paddb      xmm0, xmm5            // -> unsigned
 
     // step 3 - store 8 U and 8 V values
-    sub        ecx, 16
     movlps     qword ptr [edx], xmm0 // U
     movhps     qword ptr [edx + edi], xmm0 // V
     lea        edx, [edx + 8]
+    sub        ecx, 16
     jg         convertloop
 
     pop        edi
+    pop        esi
     ret
   }
 }
 
 __declspec(naked) __declspec(align(16))
-void BGRAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+void ABGRToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
                        uint8* dst_u, uint8* dst_v, int width) {
   __asm {
     push       esi
@@ -1795,22 +1460,26 @@ void BGRAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
     mov        edx, [esp + 8 + 12]  // dst_u
     mov        edi, [esp + 8 + 16]  // dst_v
     mov        ecx, [esp + 8 + 20]  // pix
-    movdqa     xmm7, kBGRAToU
-    movdqa     xmm6, kBGRAToV
     movdqa     xmm5, kAddUV128
+    movdqa     xmm6, kABGRToV
+    movdqa     xmm7, kABGRToU
     sub        edi, edx             // stride from u to v
 
-    align      4
  convertloop:
     /* step 1 - subsample 16x2 argb pixels to 8x1 */
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    movdqa     xmm2, [eax + 32]
-    movdqa     xmm3, [eax + 48]
-    pavgb      xmm0, [eax + esi]
-    pavgb      xmm1, [eax + esi + 16]
-    pavgb      xmm2, [eax + esi + 32]
-    pavgb      xmm3, [eax + esi + 48]
+    movdqu     xmm0, [eax]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+
     lea        eax,  [eax + 64]
     movdqa     xmm4, xmm0
     shufps     xmm0, xmm1, 0x88
@@ -1838,10 +1507,10 @@ void BGRAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
     paddb      xmm0, xmm5            // -> unsigned
 
     // step 3 - store 8 U and 8 V values
-    sub        ecx, 16
     movlps     qword ptr [edx], xmm0 // U
     movhps     qword ptr [edx + edi], xmm0 // V
     lea        edx, [edx + 8]
+    sub        ecx, 16
     jg         convertloop
 
     pop        edi
@@ -1851,8 +1520,8 @@ void BGRAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
 }
 
 __declspec(naked) __declspec(align(16))
-void BGRAToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                                 uint8* dst_u, uint8* dst_v, int width) {
+void RGBAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
   __asm {
     push       esi
     push       edi
@@ -1861,26 +1530,26 @@ void BGRAToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
     mov        edx, [esp + 8 + 12]  // dst_u
     mov        edi, [esp + 8 + 16]  // dst_v
     mov        ecx, [esp + 8 + 20]  // pix
-    movdqa     xmm7, kBGRAToU
-    movdqa     xmm6, kBGRAToV
     movdqa     xmm5, kAddUV128
+    movdqa     xmm6, kRGBAToV
+    movdqa     xmm7, kRGBAToU
     sub        edi, edx             // stride from u to v
 
-    align      4
  convertloop:
     /* step 1 - subsample 16x2 argb pixels to 8x1 */
     movdqu     xmm0, [eax]
-    movdqu     xmm1, [eax + 16]
-    movdqu     xmm2, [eax + 32]
-    movdqu     xmm3, [eax + 48]
     movdqu     xmm4, [eax + esi]
     pavgb      xmm0, xmm4
+    movdqu     xmm1, [eax + 16]
     movdqu     xmm4, [eax + esi + 16]
     pavgb      xmm1, xmm4
+    movdqu     xmm2, [eax + 32]
     movdqu     xmm4, [eax + esi + 32]
     pavgb      xmm2, xmm4
+    movdqu     xmm3, [eax + 48]
     movdqu     xmm4, [eax + esi + 48]
     pavgb      xmm3, xmm4
+
     lea        eax,  [eax + 64]
     movdqa     xmm4, xmm0
     shufps     xmm0, xmm1, 0x88
@@ -1908,10 +1577,10 @@ void BGRAToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
     paddb      xmm0, xmm5            // -> unsigned
 
     // step 3 - store 8 U and 8 V values
-    sub        ecx, 16
     movlps     qword ptr [edx], xmm0 // U
     movhps     qword ptr [edx + edi], xmm0 // V
     lea        edx, [edx + 8]
+    sub        ecx, 16
     jg         convertloop
 
     pop        edi
@@ -1919,314 +1588,259 @@ void BGRAToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
     ret
   }
 }
+#endif  // HAS_ARGBTOYROW_SSSE3
 
+// Read 8 UV from 422, upsample to 16 UV.
+#define READYUV422_AVX2 __asm {                                                \
+    __asm vmovq      xmm0, qword ptr [esi]        /* U */         /* NOLINT */ \
+    __asm vmovq      xmm1, qword ptr [esi + edi]  /* V */         /* NOLINT */ \
+    __asm lea        esi,  [esi + 8]                                           \
+    __asm vpunpcklbw ymm0, ymm0, ymm1             /* UV */                     \
+    __asm vpermq     ymm0, ymm0, 0xd8                                          \
+    __asm vpunpcklwd ymm0, ymm0, ymm0             /* UVUV (upsample) */        \
+  }
+
+// Read 8 UV from NV12, upsample to 16 UV.
+#define READNV12_AVX2 __asm {                                                  \
+    __asm vmovdqu    xmm0, [esi]                  /* UV */                     \
+    __asm lea        esi,  [esi + 16]                                          \
+    __asm vpermq     ymm0, ymm0, 0xd8                                          \
+    __asm vpunpcklwd ymm0, ymm0, ymm0             /* UVUV (upsample) */        \
+  }
+
+// Convert 16 pixels: 16 UV and 16 Y.
+#define YUVTORGB_AVX2(YuvConstants) __asm {                                    \
+    /* Step 1: Find 8 UV contributions to 16 R,G,B values */                   \
+    __asm vpmaddubsw ymm2, ymm0, YuvConstants.kUVToR        /* scale R UV */   \
+    __asm vpmaddubsw ymm1, ymm0, YuvConstants.kUVToG        /* scale G UV */   \
+    __asm vpmaddubsw ymm0, ymm0, YuvConstants.kUVToB        /* scale B UV */   \
+    __asm vmovdqu    ymm3, YuvConstants.kUVBiasR                               \
+    __asm vpsubw     ymm2, ymm3, ymm2                                          \
+    __asm vmovdqu    ymm3, YuvConstants.kUVBiasG                               \
+    __asm vpsubw     ymm1, ymm3, ymm1                                          \
+    __asm vmovdqu    ymm3, YuvConstants.kUVBiasB                               \
+    __asm vpsubw     ymm0, ymm3, ymm0                                          \
+    /* Step 2: Find Y contribution to 16 R,G,B values */                       \
+    __asm vmovdqu    xmm3, [eax]                  /* NOLINT */                 \
+    __asm lea        eax, [eax + 16]                                           \
+    __asm vpermq     ymm3, ymm3, 0xd8                                          \
+    __asm vpunpcklbw ymm3, ymm3, ymm3                                          \
+    __asm vpmulhuw   ymm3, ymm3, YuvConstants.kYToRgb                          \
+    __asm vpaddsw    ymm0, ymm0, ymm3           /* B += Y */                   \
+    __asm vpaddsw    ymm1, ymm1, ymm3           /* G += Y */                   \
+    __asm vpaddsw    ymm2, ymm2, ymm3           /* R += Y */                   \
+    __asm vpsraw     ymm0, ymm0, 6                                             \
+    __asm vpsraw     ymm1, ymm1, 6                                             \
+    __asm vpsraw     ymm2, ymm2, 6                                             \
+    __asm vpackuswb  ymm0, ymm0, ymm0           /* B */                        \
+    __asm vpackuswb  ymm1, ymm1, ymm1           /* G */                        \
+    __asm vpackuswb  ymm2, ymm2, ymm2           /* R */                        \
+  }
+
+// Store 16 ARGB values.
+#define STOREARGB_AVX2 __asm {                                                 \
+    /* Step 3: Weave into ARGB */                                              \
+    __asm vpunpcklbw ymm0, ymm0, ymm1           /* BG */                       \
+    __asm vpermq     ymm0, ymm0, 0xd8                                          \
+    __asm vpunpcklbw ymm2, ymm2, ymm5           /* RA */                       \
+    __asm vpermq     ymm2, ymm2, 0xd8                                          \
+    __asm vpunpcklwd ymm1, ymm0, ymm2           /* BGRA first 8 pixels */      \
+    __asm vpunpckhwd ymm0, ymm0, ymm2           /* BGRA next 8 pixels */       \
+    __asm vmovdqu    [edx], ymm1                                               \
+    __asm vmovdqu    [edx + 32], ymm0                                          \
+    __asm lea        edx,  [edx + 64]                                          \
+  }
+
+#ifdef HAS_I422TOARGBROW_AVX2
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
 __declspec(naked) __declspec(align(16))
-void ABGRToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                       uint8* dst_u, uint8* dst_v, int width) {
+void I422ToARGBRow_AVX2(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* dst_argb,
+                        int width) {
   __asm {
     push       esi
     push       edi
-    mov        eax, [esp + 8 + 4]   // src_argb
-    mov        esi, [esp + 8 + 8]   // src_stride_argb
-    mov        edx, [esp + 8 + 12]  // dst_u
-    mov        edi, [esp + 8 + 16]  // dst_v
-    mov        ecx, [esp + 8 + 20]  // pix
-    movdqa     xmm7, kABGRToU
-    movdqa     xmm6, kABGRToV
-    movdqa     xmm5, kAddUV128
-    sub        edi, edx             // stride from u to v
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
 
-    align      4
  convertloop:
-    /* step 1 - subsample 16x2 argb pixels to 8x1 */
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    movdqa     xmm2, [eax + 32]
-    movdqa     xmm3, [eax + 48]
-    pavgb      xmm0, [eax + esi]
-    pavgb      xmm1, [eax + esi + 16]
-    pavgb      xmm2, [eax + esi + 32]
-    pavgb      xmm3, [eax + esi + 48]
-    lea        eax,  [eax + 64]
-    movdqa     xmm4, xmm0
-    shufps     xmm0, xmm1, 0x88
-    shufps     xmm4, xmm1, 0xdd
-    pavgb      xmm0, xmm4
-    movdqa     xmm4, xmm2
-    shufps     xmm2, xmm3, 0x88
-    shufps     xmm4, xmm3, 0xdd
-    pavgb      xmm2, xmm4
-
-    // step 2 - convert to U and V
-    // from here down is very similar to Y code except
-    // instead of 16 different pixels, its 8 pixels of U and 8 of V
-    movdqa     xmm1, xmm0
-    movdqa     xmm3, xmm2
-    pmaddubsw  xmm0, xmm7  // U
-    pmaddubsw  xmm2, xmm7
-    pmaddubsw  xmm1, xmm6  // V
-    pmaddubsw  xmm3, xmm6
-    phaddw     xmm0, xmm2
-    phaddw     xmm1, xmm3
-    psraw      xmm0, 8
-    psraw      xmm1, 8
-    packsswb   xmm0, xmm1
-    paddb      xmm0, xmm5            // -> unsigned
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+    STOREARGB_AVX2
 
-    // step 3 - store 8 U and 8 V values
     sub        ecx, 16
-    movlps     qword ptr [edx], xmm0 // U
-    movhps     qword ptr [edx + edi], xmm0 // V
-    lea        edx, [edx + 8]
     jg         convertloop
 
     pop        edi
     pop        esi
+    vzeroupper
     ret
   }
 }
+#endif  // HAS_I422TOARGBROW_AVX2
 
+// 16 pixels.
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
 __declspec(naked) __declspec(align(16))
-void ABGRToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                                 uint8* dst_u, uint8* dst_v, int width) {
+void NV12ToARGBRow_AVX2(const uint8* y_buf,
+                        const uint8* uv_buf,
+                        uint8* dst_argb,
+                        int width) {
   __asm {
     push       esi
-    push       edi
-    mov        eax, [esp + 8 + 4]   // src_argb
-    mov        esi, [esp + 8 + 8]   // src_stride_argb
-    mov        edx, [esp + 8 + 12]  // dst_u
-    mov        edi, [esp + 8 + 16]  // dst_v
-    mov        ecx, [esp + 8 + 20]  // pix
-    movdqa     xmm7, kABGRToU
-    movdqa     xmm6, kABGRToV
-    movdqa     xmm5, kAddUV128
-    sub        edi, edx             // stride from u to v
+    mov        eax, [esp + 4 + 4]   // Y
+    mov        esi, [esp + 4 + 8]   // UV
+    mov        edx, [esp + 4 + 12]  // argb
+    mov        ecx, [esp + 4 + 16]  // width
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
 
-    align      4
  convertloop:
-    /* step 1 - subsample 16x2 argb pixels to 8x1 */
-    movdqu     xmm0, [eax]
-    movdqu     xmm1, [eax + 16]
-    movdqu     xmm2, [eax + 32]
-    movdqu     xmm3, [eax + 48]
-    movdqu     xmm4, [eax + esi]
-    pavgb      xmm0, xmm4
-    movdqu     xmm4, [eax + esi + 16]
-    pavgb      xmm1, xmm4
-    movdqu     xmm4, [eax + esi + 32]
-    pavgb      xmm2, xmm4
-    movdqu     xmm4, [eax + esi + 48]
-    pavgb      xmm3, xmm4
-    lea        eax,  [eax + 64]
-    movdqa     xmm4, xmm0
-    shufps     xmm0, xmm1, 0x88
-    shufps     xmm4, xmm1, 0xdd
-    pavgb      xmm0, xmm4
-    movdqa     xmm4, xmm2
-    shufps     xmm2, xmm3, 0x88
-    shufps     xmm4, xmm3, 0xdd
-    pavgb      xmm2, xmm4
+    READNV12_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+    STOREARGB_AVX2
 
-    // step 2 - convert to U and V
-    // from here down is very similar to Y code except
-    // instead of 16 different pixels, its 8 pixels of U and 8 of V
-    movdqa     xmm1, xmm0
-    movdqa     xmm3, xmm2
-    pmaddubsw  xmm0, xmm7  // U
-    pmaddubsw  xmm2, xmm7
-    pmaddubsw  xmm1, xmm6  // V
-    pmaddubsw  xmm3, xmm6
-    phaddw     xmm0, xmm2
-    phaddw     xmm1, xmm3
-    psraw      xmm0, 8
-    psraw      xmm1, 8
-    packsswb   xmm0, xmm1
-    paddb      xmm0, xmm5            // -> unsigned
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+
+// 16 pixels.
+// 8 VU values upsampled to 16 VU, mixed with 16 Y producing 16 ARGB (64 bytes).
+__declspec(naked) __declspec(align(16))
+void NV21ToARGBRow_AVX2(const uint8* y_buf,
+                        const uint8* uv_buf,
+                        uint8* dst_argb,
+                        int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // Y
+    mov        esi, [esp + 4 + 8]   // UV
+    mov        edx, [esp + 4 + 12]  // argb
+    mov        ecx, [esp + 4 + 16]  // width
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+
+ convertloop:
+    READNV12_AVX2
+    YUVTORGB_AVX2(kYvuConstants)
+    STOREARGB_AVX2
 
-    // step 3 - store 8 U and 8 V values
     sub        ecx, 16
-    movlps     qword ptr [edx], xmm0 // U
-    movhps     qword ptr [edx + edi], xmm0 // V
-    lea        edx, [edx + 8]
     jg         convertloop
 
-    pop        edi
     pop        esi
     ret
   }
 }
 
+#ifdef HAS_I422TOBGRAROW_AVX2
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 BGRA (64 bytes).
+// TODO(fbarchard): Use macros to reduce duplicate code.  See SSSE3.
 __declspec(naked) __declspec(align(16))
-void RGBAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                       uint8* dst_u, uint8* dst_v, int width) {
+void I422ToBGRARow_AVX2(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* dst_argb,
+                        int width) {
   __asm {
     push       esi
     push       edi
-    mov        eax, [esp + 8 + 4]   // src_argb
-    mov        esi, [esp + 8 + 8]   // src_stride_argb
-    mov        edx, [esp + 8 + 12]  // dst_u
-    mov        edi, [esp + 8 + 16]  // dst_v
-    mov        ecx, [esp + 8 + 20]  // pix
-    movdqa     xmm7, kRGBAToU
-    movdqa     xmm6, kRGBAToV
-    movdqa     xmm5, kAddUV128
-    sub        edi, edx             // stride from u to v
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
 
-    align      4
  convertloop:
-    /* step 1 - subsample 16x2 argb pixels to 8x1 */
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    movdqa     xmm2, [eax + 32]
-    movdqa     xmm3, [eax + 48]
-    pavgb      xmm0, [eax + esi]
-    pavgb      xmm1, [eax + esi + 16]
-    pavgb      xmm2, [eax + esi + 32]
-    pavgb      xmm3, [eax + esi + 48]
-    lea        eax,  [eax + 64]
-    movdqa     xmm4, xmm0
-    shufps     xmm0, xmm1, 0x88
-    shufps     xmm4, xmm1, 0xdd
-    pavgb      xmm0, xmm4
-    movdqa     xmm4, xmm2
-    shufps     xmm2, xmm3, 0x88
-    shufps     xmm4, xmm3, 0xdd
-    pavgb      xmm2, xmm4
-
-    // step 2 - convert to U and V
-    // from here down is very similar to Y code except
-    // instead of 16 different pixels, its 8 pixels of U and 8 of V
-    movdqa     xmm1, xmm0
-    movdqa     xmm3, xmm2
-    pmaddubsw  xmm0, xmm7  // U
-    pmaddubsw  xmm2, xmm7
-    pmaddubsw  xmm1, xmm6  // V
-    pmaddubsw  xmm3, xmm6
-    phaddw     xmm0, xmm2
-    phaddw     xmm1, xmm3
-    psraw      xmm0, 8
-    psraw      xmm1, 8
-    packsswb   xmm0, xmm1
-    paddb      xmm0, xmm5            // -> unsigned
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
 
-    // step 3 - store 8 U and 8 V values
+    // Step 3: Weave into BGRA
+    vpunpcklbw ymm1, ymm1, ymm0           // GB
+    vpermq     ymm1, ymm1, 0xd8
+    vpunpcklbw ymm2, ymm5, ymm2           // AR
+    vpermq     ymm2, ymm2, 0xd8
+    vpunpcklwd ymm0, ymm2, ymm1           // ARGB first 8 pixels
+    vpunpckhwd ymm2, ymm2, ymm1           // ARGB next 8 pixels
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + 32], ymm2
+    lea        edx,  [edx + 64]
     sub        ecx, 16
-    movlps     qword ptr [edx], xmm0 // U
-    movhps     qword ptr [edx + edi], xmm0 // V
-    lea        edx, [edx + 8]
     jg         convertloop
 
     pop        edi
     pop        esi
+    vzeroupper
     ret
   }
 }
+#endif  // HAS_I422TOBGRAROW_AVX2
 
+#ifdef HAS_I422TORGBAROW_AVX2
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 RGBA (64 bytes).
+// TODO(fbarchard): Use macros to reduce duplicate code.  See SSSE3.
 __declspec(naked) __declspec(align(16))
-void RGBAToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                                 uint8* dst_u, uint8* dst_v, int width) {
+void I422ToRGBARow_AVX2(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* dst_argb,
+                        int width) {
   __asm {
     push       esi
     push       edi
-    mov        eax, [esp + 8 + 4]   // src_argb
-    mov        esi, [esp + 8 + 8]   // src_stride_argb
-    mov        edx, [esp + 8 + 12]  // dst_u
-    mov        edi, [esp + 8 + 16]  // dst_v
-    mov        ecx, [esp + 8 + 20]  // pix
-    movdqa     xmm7, kRGBAToU
-    movdqa     xmm6, kRGBAToV
-    movdqa     xmm5, kAddUV128
-    sub        edi, edx             // stride from u to v
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
 
-    align      4
  convertloop:
-    /* step 1 - subsample 16x2 argb pixels to 8x1 */
-    movdqu     xmm0, [eax]
-    movdqu     xmm1, [eax + 16]
-    movdqu     xmm2, [eax + 32]
-    movdqu     xmm3, [eax + 48]
-    movdqu     xmm4, [eax + esi]
-    pavgb      xmm0, xmm4
-    movdqu     xmm4, [eax + esi + 16]
-    pavgb      xmm1, xmm4
-    movdqu     xmm4, [eax + esi + 32]
-    pavgb      xmm2, xmm4
-    movdqu     xmm4, [eax + esi + 48]
-    pavgb      xmm3, xmm4
-    lea        eax,  [eax + 64]
-    movdqa     xmm4, xmm0
-    shufps     xmm0, xmm1, 0x88
-    shufps     xmm4, xmm1, 0xdd
-    pavgb      xmm0, xmm4
-    movdqa     xmm4, xmm2
-    shufps     xmm2, xmm3, 0x88
-    shufps     xmm4, xmm3, 0xdd
-    pavgb      xmm2, xmm4
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
 
-    // step 2 - convert to U and V
-    // from here down is very similar to Y code except
-    // instead of 16 different pixels, its 8 pixels of U and 8 of V
-    movdqa     xmm1, xmm0
-    movdqa     xmm3, xmm2
-    pmaddubsw  xmm0, xmm7  // U
-    pmaddubsw  xmm2, xmm7
-    pmaddubsw  xmm1, xmm6  // V
-    pmaddubsw  xmm3, xmm6
-    phaddw     xmm0, xmm2
-    phaddw     xmm1, xmm3
-    psraw      xmm0, 8
-    psraw      xmm1, 8
-    packsswb   xmm0, xmm1
-    paddb      xmm0, xmm5            // -> unsigned
-
-    // step 3 - store 8 U and 8 V values
+    // Step 3: Weave into RGBA
+    vpunpcklbw ymm1, ymm1, ymm2           // GR
+    vpermq     ymm1, ymm1, 0xd8
+    vpunpcklbw ymm2, ymm5, ymm0           // AB
+    vpermq     ymm2, ymm2, 0xd8
+    vpunpcklwd ymm0, ymm2, ymm1           // ABGR first 8 pixels
+    vpunpckhwd ymm1, ymm2, ymm1           // ABGR next 8 pixels
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + 32], ymm1
+    lea        edx,  [edx + 64]
     sub        ecx, 16
-    movlps     qword ptr [edx], xmm0 // U
-    movhps     qword ptr [edx + edi], xmm0 // V
-    lea        edx, [edx + 8]
     jg         convertloop
 
     pop        edi
     pop        esi
+    vzeroupper
     ret
   }
 }
-#endif  // HAS_ARGBTOYROW_SSSE3
-
-#ifdef HAS_I422TOARGBROW_AVX2
-
-static const lvec8 kUVToB_AVX = {
-  UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB,
-  UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB
-};
-static const lvec8 kUVToR_AVX = {
-  UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR,
-  UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR
-};
-static const lvec8 kUVToG_AVX = {
-  UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG,
-  UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG
-};
-static const lvec16 kYToRgb_AVX = {
-  YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG
-};
-static const lvec16 kYSub16_AVX = {
-  16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16
-};
-static const lvec16 kUVBiasB_AVX = {
-  BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB
-};
-static const lvec16 kUVBiasG_AVX = {
-  BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG
-};
-static const lvec16 kUVBiasR_AVX = {
-  BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR
-};
+#endif  // HAS_I422TORGBAROW_AVX2
 
+#ifdef HAS_I422TOABGRROW_AVX2
 // 16 pixels
-// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ABGR (64 bytes).
+// TODO(fbarchard): Use macros to reduce duplicate code.  See SSSE3.
 __declspec(naked) __declspec(align(16))
-void I422ToARGBRow_AVX2(const uint8* y_buf,
+void I422ToABGRRow_AVX2(const uint8* y_buf,
                         const uint8* u_buf,
                         const uint8* v_buf,
                         uint8* dst_argb,
@@ -2241,63 +1855,33 @@ void I422ToARGBRow_AVX2(const uint8* y_buf,
     mov        ecx, [esp + 8 + 20]  // width
     sub        edi, esi
     vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
-    vpxor      ymm4, ymm4, ymm4
 
-    align      4
  convertloop:
-    vmovq      xmm0, qword ptr [esi]          //  U
-    vmovq      xmm1, qword ptr [esi + edi]    //  V
-    lea        esi,  [esi + 8]
-    vpunpcklbw ymm0, ymm0, ymm1               // UV
-    vpermq     ymm0, ymm0, 0xd8
-    vpunpcklwd ymm0, ymm0, ymm0              // UVUV
-    vpmaddubsw ymm2, ymm0, kUVToB_AVX        // scale B UV
-    vpmaddubsw ymm1, ymm0, kUVToG_AVX        // scale G UV
-    vpmaddubsw ymm0, ymm0, kUVToR_AVX        // scale R UV
-    vpsubw     ymm2, ymm2, kUVBiasB_AVX      // unbias back to signed
-    vpsubw     ymm1, ymm1, kUVBiasG_AVX
-    vpsubw     ymm0, ymm0, kUVBiasR_AVX
-
-    // Step 2: Find Y contribution to 16 R,G,B values
-    vmovdqu    xmm3, [eax]                  // NOLINT
-    lea        eax, [eax + 16]
-    vpermq     ymm3, ymm3, 0xd8
-    vpunpcklbw ymm3, ymm3, ymm4
-    vpsubsw    ymm3, ymm3, kYSub16_AVX
-    vpmullw    ymm3, ymm3, kYToRgb_AVX
-    vpaddsw    ymm2, ymm2, ymm3           // B += Y
-    vpaddsw    ymm1, ymm1, ymm3           // G += Y
-    vpaddsw    ymm0, ymm0, ymm3           // R += Y
-    vpsraw     ymm2, ymm2, 6
-    vpsraw     ymm1, ymm1, 6
-    vpsraw     ymm0, ymm0, 6
-    vpackuswb  ymm2, ymm2, ymm2           // B
-    vpackuswb  ymm1, ymm1, ymm1           // G
-    vpackuswb  ymm0, ymm0, ymm0           // R
-
-    // Step 3: Weave into ARGB
-    vpunpcklbw ymm2, ymm2, ymm1           // BG
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+
+    // Step 3: Weave into ABGR
+    vpunpcklbw ymm1, ymm2, ymm1           // RG
+    vpermq     ymm1, ymm1, 0xd8
+    vpunpcklbw ymm2, ymm0, ymm5           // BA
     vpermq     ymm2, ymm2, 0xd8
-    vpunpcklbw ymm0, ymm0, ymm5           // RA
-    vpermq     ymm0, ymm0, 0xd8
-    vpunpcklwd ymm1, ymm2, ymm0           // BGRA first 8 pixels
-    vpunpckhwd ymm2, ymm2, ymm0           // BGRA next 8 pixels
-    vmovdqu    [edx], ymm1
-    vmovdqu    [edx + 32], ymm2
+    vpunpcklwd ymm0, ymm1, ymm2           // RGBA first 8 pixels
+    vpunpckhwd ymm1, ymm1, ymm2           // RGBA next 8 pixels
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + 32], ymm1
     lea        edx,  [edx + 64]
     sub        ecx, 16
     jg         convertloop
-    vzeroupper
 
     pop        edi
     pop        esi
+    vzeroupper
     ret
   }
 }
-#endif  // HAS_I422TOARGBROW_AVX2
-
-#ifdef HAS_I422TOARGBROW_SSSE3
+#endif  // HAS_I422TOABGRROW_AVX2
 
+#if defined(HAS_I422TOARGBROW_SSSE3)
 // TODO(fbarchard): Read that does half size on Y and treats 420 as 444.
 
 // Read 8 UV from 444.
@@ -2337,22 +1921,25 @@ void I422ToARGBRow_AVX2(const uint8* y_buf,
   }
 
 // Convert 8 pixels: 8 UV and 8 Y.
-#define YUVTORGB __asm {                                                       \
+#define YUVTORGB(YuvConstants) __asm {                                         \
     /* Step 1: Find 4 UV contributions to 8 R,G,B values */                    \
     __asm movdqa     xmm1, xmm0                                                \
     __asm movdqa     xmm2, xmm0                                                \
-    __asm pmaddubsw  xmm0, kUVToB        /* scale B UV */                      \
-    __asm pmaddubsw  xmm1, kUVToG        /* scale G UV */                      \
-    __asm pmaddubsw  xmm2, kUVToR        /* scale R UV */                      \
-    __asm psubw      xmm0, kUVBiasB      /* unbias back to signed */           \
-    __asm psubw      xmm1, kUVBiasG                                            \
-    __asm psubw      xmm2, kUVBiasR                                            \
+    __asm movdqa     xmm3, xmm0                                                \
+    __asm movdqa     xmm0, YuvConstants.kUVBiasB /* unbias back to signed */   \
+    __asm pmaddubsw  xmm1, YuvConstants.kUVToB   /* scale B UV */              \
+    __asm psubw      xmm0, xmm1                                                \
+    __asm movdqa     xmm1, YuvConstants.kUVBiasG                               \
+    __asm pmaddubsw  xmm2, YuvConstants.kUVToG   /* scale G UV */              \
+    __asm psubw      xmm1, xmm2                                                \
+    __asm movdqa     xmm2, YuvConstants.kUVBiasR                               \
+    __asm pmaddubsw  xmm3, YuvConstants.kUVToR   /* scale R UV */              \
+    __asm psubw      xmm2, xmm3                                                \
     /* Step 2: Find Y contribution to 8 R,G,B values */                        \
     __asm movq       xmm3, qword ptr [eax]                        /* NOLINT */ \
     __asm lea        eax, [eax + 8]                                            \
-    __asm punpcklbw  xmm3, xmm4                                                \
-    __asm psubsw     xmm3, kYSub16                                             \
-    __asm pmullw     xmm3, kYToRgb                                             \
+    __asm punpcklbw  xmm3, xmm3                                                \
+    __asm pmulhuw    xmm3, YuvConstants.kYToRgb                                \
     __asm paddsw     xmm0, xmm3           /* B += Y */                         \
     __asm paddsw     xmm1, xmm3           /* G += Y */                         \
     __asm paddsw     xmm2, xmm3           /* R += Y */                         \
@@ -2364,35 +1951,131 @@ void I422ToARGBRow_AVX2(const uint8* y_buf,
     __asm packuswb   xmm2, xmm2           /* R */                              \
   }
 
-// Convert 8 pixels: 8 VU and 8 Y.
-#define YVUTORGB __asm {                                                       \
-    /* Step 1: Find 4 UV contributions to 8 R,G,B values */                    \
+// Store 8 ARGB values.
+#define STOREARGB __asm {                                                      \
+    /* Step 3: Weave into ARGB */                                              \
+    __asm punpcklbw  xmm0, xmm1           /* BG */                             \
+    __asm punpcklbw  xmm2, xmm5           /* RA */                             \
     __asm movdqa     xmm1, xmm0                                                \
-    __asm movdqa     xmm2, xmm0                                                \
-    __asm pmaddubsw  xmm0, kVUToB        /* scale B UV */                      \
-    __asm pmaddubsw  xmm1, kVUToG        /* scale G UV */                      \
-    __asm pmaddubsw  xmm2, kVUToR        /* scale R UV */                      \
-    __asm psubw      xmm0, kUVBiasB      /* unbias back to signed */           \
-    __asm psubw      xmm1, kUVBiasG                                            \
-    __asm psubw      xmm2, kUVBiasR                                            \
-    /* Step 2: Find Y contribution to 8 R,G,B values */                        \
-    __asm movq       xmm3, qword ptr [eax]                        /* NOLINT */ \
-    __asm lea        eax, [eax + 8]                                            \
-    __asm punpcklbw  xmm3, xmm4                                                \
-    __asm psubsw     xmm3, kYSub16                                             \
-    __asm pmullw     xmm3, kYToRgb                                             \
-    __asm paddsw     xmm0, xmm3           /* B += Y */                         \
-    __asm paddsw     xmm1, xmm3           /* G += Y */                         \
-    __asm paddsw     xmm2, xmm3           /* R += Y */                         \
-    __asm psraw      xmm0, 6                                                   \
-    __asm psraw      xmm1, 6                                                   \
-    __asm psraw      xmm2, 6                                                   \
-    __asm packuswb   xmm0, xmm0           /* B */                              \
-    __asm packuswb   xmm1, xmm1           /* G */                              \
-    __asm packuswb   xmm2, xmm2           /* R */                              \
-  }
-
-// 8 pixels, dest aligned 16.
+    __asm punpcklwd  xmm0, xmm2           /* BGRA first 4 pixels */            \
+    __asm punpckhwd  xmm1, xmm2           /* BGRA next 4 pixels */             \
+    __asm movdqu     [edx], xmm0                                               \
+    __asm movdqu     [edx + 16], xmm1                                          \
+    __asm lea        edx,  [edx + 32]                                          \
+  }
+
+// Store 8 BGRA values.
+#define STOREBGRA __asm {                                                      \
+    /* Step 3: Weave into BGRA */                                              \
+    __asm pcmpeqb    xmm5, xmm5           /* generate 0xffffffff for alpha */  \
+    __asm punpcklbw  xmm1, xmm0           /* GB */                             \
+    __asm punpcklbw  xmm5, xmm2           /* AR */                             \
+    __asm movdqa     xmm0, xmm5                                                \
+    __asm punpcklwd  xmm5, xmm1           /* BGRA first 4 pixels */            \
+    __asm punpckhwd  xmm0, xmm1           /* BGRA next 4 pixels */             \
+    __asm movdqu     [edx], xmm5                                               \
+    __asm movdqu     [edx + 16], xmm0                                          \
+    __asm lea        edx,  [edx + 32]                                          \
+  }
+
+// Store 8 ABGR values.
+#define STOREABGR __asm {                                                      \
+    /* Step 3: Weave into ABGR */                                              \
+    __asm punpcklbw  xmm2, xmm1           /* RG */                             \
+    __asm punpcklbw  xmm0, xmm5           /* BA */                             \
+    __asm movdqa     xmm1, xmm2                                                \
+    __asm punpcklwd  xmm2, xmm0           /* RGBA first 4 pixels */            \
+    __asm punpckhwd  xmm1, xmm0           /* RGBA next 4 pixels */             \
+    __asm movdqu     [edx], xmm2                                               \
+    __asm movdqu     [edx + 16], xmm1                                          \
+    __asm lea        edx,  [edx + 32]                                          \
+  }
+
+// Store 8 RGBA values.
+#define STORERGBA __asm {                                                      \
+    /* Step 3: Weave into RGBA */                                              \
+    __asm pcmpeqb    xmm5, xmm5           /* generate 0xffffffff for alpha */  \
+    __asm punpcklbw  xmm1, xmm2           /* GR */                             \
+    __asm punpcklbw  xmm5, xmm0           /* AB */                             \
+    __asm movdqa     xmm0, xmm5                                                \
+    __asm punpcklwd  xmm5, xmm1           /* RGBA first 4 pixels */            \
+    __asm punpckhwd  xmm0, xmm1           /* RGBA next 4 pixels */             \
+    __asm movdqu     [edx], xmm5                                               \
+    __asm movdqu     [edx + 16], xmm0                                          \
+    __asm lea        edx,  [edx + 32]                                          \
+  }
+
+// Store 8 RGB24 values.
+#define STORERGB24 __asm {                                                     \
+    /* Step 3: Weave into RRGB */                                              \
+    __asm punpcklbw  xmm0, xmm1           /* BG */                             \
+    __asm punpcklbw  xmm2, xmm2           /* RR */                             \
+    __asm movdqa     xmm1, xmm0                                                \
+    __asm punpcklwd  xmm0, xmm2           /* BGRR first 4 pixels */            \
+    __asm punpckhwd  xmm1, xmm2           /* BGRR next 4 pixels */             \
+    /* Step 4: RRGB -> RGB24 */                                                \
+    __asm pshufb     xmm0, xmm5           /* Pack first 8 and last 4 bytes. */ \
+    __asm pshufb     xmm1, xmm6           /* Pack first 12 bytes. */           \
+    __asm palignr    xmm1, xmm0, 12       /* last 4 bytes of xmm0 + 12 xmm1 */ \
+    __asm movq       qword ptr [edx], xmm0  /* First 8 bytes */                \
+    __asm movdqu     [edx + 8], xmm1      /* Last 16 bytes */                  \
+    __asm lea        edx,  [edx + 24]                                          \
+  }
+
+// Store 8 RAW values.
+#define STORERAW __asm {                                                       \
+    /* Step 3: Weave into RRGB */                                              \
+    __asm punpcklbw  xmm0, xmm1           /* BG */                             \
+    __asm punpcklbw  xmm2, xmm2           /* RR */                             \
+    __asm movdqa     xmm1, xmm0                                                \
+    __asm punpcklwd  xmm0, xmm2           /* BGRR first 4 pixels */            \
+    __asm punpckhwd  xmm1, xmm2           /* BGRR next 4 pixels */             \
+    /* Step 4: RRGB -> RAW */                                                  \
+    __asm pshufb     xmm0, xmm5           /* Pack first 8 and last 4 bytes. */ \
+    __asm pshufb     xmm1, xmm6           /* Pack first 12 bytes. */           \
+    __asm palignr    xmm1, xmm0, 12       /* last 4 bytes of xmm0 + 12 xmm1 */ \
+    __asm movq       qword ptr [edx], xmm0  /* First 8 bytes */                \
+    __asm movdqu     [edx + 8], xmm1      /* Last 16 bytes */                  \
+    __asm lea        edx,  [edx + 24]                                          \
+  }
+
+// Store 8 RGB565 values.
+#define STORERGB565 __asm {                                                    \
+    /* Step 3: Weave into RRGB */                                              \
+    __asm punpcklbw  xmm0, xmm1           /* BG */                             \
+    __asm punpcklbw  xmm2, xmm2           /* RR */                             \
+    __asm movdqa     xmm1, xmm0                                                \
+    __asm punpcklwd  xmm0, xmm2           /* BGRR first 4 pixels */            \
+    __asm punpckhwd  xmm1, xmm2           /* BGRR next 4 pixels */             \
+    /* Step 4: RRGB -> RGB565 */                                               \
+    __asm movdqa     xmm3, xmm0    /* B  first 4 pixels of argb */             \
+    __asm movdqa     xmm2, xmm0    /* G */                                     \
+    __asm pslld      xmm0, 8       /* R */                                     \
+    __asm psrld      xmm3, 3       /* B */                                     \
+    __asm psrld      xmm2, 5       /* G */                                     \
+    __asm psrad      xmm0, 16      /* R */                                     \
+    __asm pand       xmm3, xmm5    /* B */                                     \
+    __asm pand       xmm2, xmm6    /* G */                                     \
+    __asm pand       xmm0, xmm7    /* R */                                     \
+    __asm por        xmm3, xmm2    /* BG */                                    \
+    __asm por        xmm0, xmm3    /* BGR */                                   \
+    __asm movdqa     xmm3, xmm1    /* B  next 4 pixels of argb */              \
+    __asm movdqa     xmm2, xmm1    /* G */                                     \
+    __asm pslld      xmm1, 8       /* R */                                     \
+    __asm psrld      xmm3, 3       /* B */                                     \
+    __asm psrld      xmm2, 5       /* G */                                     \
+    __asm psrad      xmm1, 16      /* R */                                     \
+    __asm pand       xmm3, xmm5    /* B */                                     \
+    __asm pand       xmm2, xmm6    /* G */                                     \
+    __asm pand       xmm1, xmm7    /* R */                                     \
+    __asm por        xmm3, xmm2    /* BG */                                    \
+    __asm por        xmm1, xmm3    /* BGR */                                   \
+    __asm packssdw   xmm0, xmm1                                                \
+    __asm movdqu     [edx], xmm0   /* store 8 pixels of RGB565 */              \
+    __asm lea        edx, [edx + 16]                                           \
+  }
+
+// 8 pixels.
 // 8 UV values, mixed with 8 Y producing 8 ARGB (32 bytes).
 __declspec(naked) __declspec(align(16))
 void I444ToARGBRow_SSSE3(const uint8* y_buf,
@@ -2410,22 +2093,12 @@ void I444ToARGBRow_SSSE3(const uint8* y_buf,
     mov        ecx, [esp + 8 + 20]  // width
     sub        edi, esi
     pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
-    pxor       xmm4, xmm4
 
-    align      4
  convertloop:
     READYUV444
-    YUVTORGB
+    YUVTORGB(kYuvConstants)
+    STOREARGB
 
-    // Step 3: Weave into ARGB
-    punpcklbw  xmm0, xmm1           // BG
-    punpcklbw  xmm2, xmm5           // RA
-    movdqa     xmm1, xmm0
-    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
-    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
-    movdqa     [edx], xmm0
-    movdqa     [edx + 16], xmm1
-    lea        edx,  [edx + 32]
     sub        ecx, 8
     jg         convertloop
 
@@ -2435,8 +2108,8 @@ void I444ToARGBRow_SSSE3(const uint8* y_buf,
   }
 }
 
-// 8 pixels, dest aligned 16.
-// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+// 8 pixels.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 RGB24 (24 bytes).
 __declspec(naked) __declspec(align(16))
 void I422ToRGB24Row_SSSE3(const uint8* y_buf,
                           const uint8* u_buf,
@@ -2452,27 +2125,14 @@ void I422ToRGB24Row_SSSE3(const uint8* y_buf,
     mov        edx, [esp + 8 + 16]  // rgb24
     mov        ecx, [esp + 8 + 20]  // width
     sub        edi, esi
-    pxor       xmm4, xmm4
     movdqa     xmm5, kShuffleMaskARGBToRGB24_0
     movdqa     xmm6, kShuffleMaskARGBToRGB24
 
-    align      4
  convertloop:
     READYUV422
-    YUVTORGB
+    YUVTORGB(kYuvConstants)
+    STORERGB24
 
-    // Step 3: Weave into RRGB
-    punpcklbw  xmm0, xmm1           // BG
-    punpcklbw  xmm2, xmm2           // RR
-    movdqa     xmm1, xmm0
-    punpcklwd  xmm0, xmm2           // BGRR first 4 pixels
-    punpckhwd  xmm1, xmm2           // BGRR next 4 pixels
-    pshufb     xmm0, xmm5           // Pack into first 8 and last 4 bytes.
-    pshufb     xmm1, xmm6           // Pack into first 12 bytes.
-    palignr    xmm1, xmm0, 12       // last 4 bytes of xmm0 + 12 from xmm1
-    movq       qword ptr [edx], xmm0  // First 8 bytes
-    movdqu     [edx + 8], xmm1      // Last 16 bytes. = 24 bytes, 8 RGB pixels.
-    lea        edx,  [edx + 24]
     sub        ecx, 8
     jg         convertloop
 
@@ -2482,8 +2142,8 @@ void I422ToRGB24Row_SSSE3(const uint8* y_buf,
   }
 }
 
-// 8 pixels, dest aligned 16.
-// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+// 8 pixels.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 RAW (24 bytes).
 __declspec(naked) __declspec(align(16))
 void I422ToRAWRow_SSSE3(const uint8* y_buf,
                         const uint8* u_buf,
@@ -2499,27 +2159,14 @@ void I422ToRAWRow_SSSE3(const uint8* y_buf,
     mov        edx, [esp + 8 + 16]  // raw
     mov        ecx, [esp + 8 + 20]  // width
     sub        edi, esi
-    pxor       xmm4, xmm4
     movdqa     xmm5, kShuffleMaskARGBToRAW_0
     movdqa     xmm6, kShuffleMaskARGBToRAW
 
-    align      4
  convertloop:
     READYUV422
-    YUVTORGB
+    YUVTORGB(kYuvConstants)
+    STORERAW
 
-    // Step 3: Weave into RRGB
-    punpcklbw  xmm0, xmm1           // BG
-    punpcklbw  xmm2, xmm2           // RR
-    movdqa     xmm1, xmm0
-    punpcklwd  xmm0, xmm2           // BGRR first 4 pixels
-    punpckhwd  xmm1, xmm2           // BGRR next 4 pixels
-    pshufb     xmm0, xmm5           // Pack into first 8 and last 4 bytes.
-    pshufb     xmm1, xmm6           // Pack into first 12 bytes.
-    palignr    xmm1, xmm0, 12       // last 4 bytes of xmm0 + 12 from xmm1
-    movq       qword ptr [edx], xmm0  // First 8 bytes
-    movdqu     [edx + 8], xmm1      // Last 16 bytes. = 24 bytes, 8 RGB pixels.
-    lea        edx,  [edx + 24]
     sub        ecx, 8
     jg         convertloop
 
@@ -2529,8 +2176,8 @@ void I422ToRAWRow_SSSE3(const uint8* y_buf,
   }
 }
 
-// 8 pixels, dest unaligned.
-// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+// 8 pixels
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 RGB565 (16 bytes).
 __declspec(naked) __declspec(align(16))
 void I422ToRGB565Row_SSSE3(const uint8* y_buf,
                            const uint8* u_buf,
@@ -2546,7 +2193,6 @@ void I422ToRGB565Row_SSSE3(const uint8* y_buf,
     mov        edx, [esp + 8 + 16]  // rgb565
     mov        ecx, [esp + 8 + 20]  // width
     sub        edi, esi
-    pxor       xmm4, xmm4
     pcmpeqb    xmm5, xmm5       // generate mask 0x0000001f
     psrld      xmm5, 27
     pcmpeqb    xmm6, xmm6       // generate mask 0x000007e0
@@ -2555,45 +2201,12 @@ void I422ToRGB565Row_SSSE3(const uint8* y_buf,
     pcmpeqb    xmm7, xmm7       // generate mask 0xfffff800
     pslld      xmm7, 11
 
-    align      4
  convertloop:
     READYUV422
-    YUVTORGB
+    YUVTORGB(kYuvConstants)
+    STORERGB565
 
-    // Step 3: Weave into RRGB
-    punpcklbw  xmm0, xmm1           // BG
-    punpcklbw  xmm2, xmm2           // RR
-    movdqa     xmm1, xmm0
-    punpcklwd  xmm0, xmm2           // BGRR first 4 pixels
-    punpckhwd  xmm1, xmm2           // BGRR next 4 pixels
-
-    // Step 3b: RRGB -> RGB565
-    movdqa     xmm3, xmm0    // B  first 4 pixels of argb
-    movdqa     xmm2, xmm0    // G
-    pslld      xmm0, 8       // R
-    psrld      xmm3, 3       // B
-    psrld      xmm2, 5       // G
-    psrad      xmm0, 16      // R
-    pand       xmm3, xmm5    // B
-    pand       xmm2, xmm6    // G
-    pand       xmm0, xmm7    // R
-    por        xmm3, xmm2    // BG
-    por        xmm0, xmm3    // BGR
-    movdqa     xmm3, xmm1    // B  next 4 pixels of argb
-    movdqa     xmm2, xmm1    // G
-    pslld      xmm1, 8       // R
-    psrld      xmm3, 3       // B
-    psrld      xmm2, 5       // G
-    psrad      xmm1, 16      // R
-    pand       xmm3, xmm5    // B
-    pand       xmm2, xmm6    // G
-    pand       xmm1, xmm7    // R
-    por        xmm3, xmm2    // BG
-    por        xmm1, xmm3    // BGR
-    packssdw   xmm0, xmm1
     sub        ecx, 8
-    movdqu     [edx], xmm0   // store 8 pixels of RGB565
-    lea        edx, [edx + 16]
     jg         convertloop
 
     pop        edi
@@ -2602,7 +2215,7 @@ void I422ToRGB565Row_SSSE3(const uint8* y_buf,
   }
 }
 
-// 8 pixels, dest aligned 16.
+// 8 pixels.
 // 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
 __declspec(naked) __declspec(align(16))
 void I422ToARGBRow_SSSE3(const uint8* y_buf,
@@ -2620,22 +2233,12 @@ void I422ToARGBRow_SSSE3(const uint8* y_buf,
     mov        ecx, [esp + 8 + 20]  // width
     sub        edi, esi
     pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
-    pxor       xmm4, xmm4
 
-    align      4
  convertloop:
     READYUV422
-    YUVTORGB
+    YUVTORGB(kYuvConstants)
+    STOREARGB
 
-    // Step 3: Weave into ARGB
-    punpcklbw  xmm0, xmm1           // BG
-    punpcklbw  xmm2, xmm5           // RA
-    movdqa     xmm1, xmm0
-    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
-    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
-    movdqa     [edx], xmm0
-    movdqa     [edx + 16], xmm1
-    lea        edx,  [edx + 32]
     sub        ecx, 8
     jg         convertloop
 
@@ -2645,7 +2248,7 @@ void I422ToARGBRow_SSSE3(const uint8* y_buf,
   }
 }
 
-// 8 pixels, dest aligned 16.
+// 8 pixels.
 // 2 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
 // Similar to I420 but duplicate UV once more.
 __declspec(naked) __declspec(align(16))
@@ -2664,23 +2267,13 @@ void I411ToARGBRow_SSSE3(const uint8* y_buf,
     mov        edx, [esp + 12 + 16]  // argb
     mov        ecx, [esp + 12 + 20]  // width
     sub        edi, esi
-    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
-    pxor       xmm4, xmm4
+    pcmpeqb    xmm5, xmm5            // generate 0xffffffff for alpha
 
-    align      4
  convertloop:
     READYUV411  // modifies EBX
-    YUVTORGB
+    YUVTORGB(kYuvConstants)
+    STOREARGB
 
-    // Step 3: Weave into ARGB
-    punpcklbw  xmm0, xmm1           // BG
-    punpcklbw  xmm2, xmm5           // RA
-    movdqa     xmm1, xmm0
-    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
-    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
-    movdqa     [edx], xmm0
-    movdqa     [edx + 16], xmm1
-    lea        edx,  [edx + 32]
     sub        ecx, 8
     jg         convertloop
 
@@ -2691,7 +2284,7 @@ void I411ToARGBRow_SSSE3(const uint8* y_buf,
   }
 }
 
-// 8 pixels, dest aligned 16.
+// 8 pixels.
 // 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
 __declspec(naked) __declspec(align(16))
 void NV12ToARGBRow_SSSE3(const uint8* y_buf,
@@ -2705,22 +2298,12 @@ void NV12ToARGBRow_SSSE3(const uint8* y_buf,
     mov        edx, [esp + 4 + 12]  // argb
     mov        ecx, [esp + 4 + 16]  // width
     pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
-    pxor       xmm4, xmm4
 
-    align      4
  convertloop:
     READNV12
-    YUVTORGB
+    YUVTORGB(kYuvConstants)
+    STOREARGB
 
-    // Step 3: Weave into ARGB
-    punpcklbw  xmm0, xmm1           // BG
-    punpcklbw  xmm2, xmm5           // RA
-    movdqa     xmm1, xmm0
-    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
-    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
-    movdqa     [edx], xmm0
-    movdqa     [edx + 16], xmm1
-    lea        edx,  [edx + 32]
     sub        ecx, 8
     jg         convertloop
 
@@ -2729,8 +2312,8 @@ void NV12ToARGBRow_SSSE3(const uint8* y_buf,
   }
 }
 
-// 8 pixels, dest aligned 16.
-// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+// 8 pixels.
+// 4 VU values upsampled to 8 VU, mixed with 8 Y producing 8 ARGB (32 bytes).
 __declspec(naked) __declspec(align(16))
 void NV21ToARGBRow_SSSE3(const uint8* y_buf,
                          const uint8* uv_buf,
@@ -2739,234 +2322,16 @@ void NV21ToARGBRow_SSSE3(const uint8* y_buf,
   __asm {
     push       esi
     mov        eax, [esp + 4 + 4]   // Y
-    mov        esi, [esp + 4 + 8]   // VU
-    mov        edx, [esp + 4 + 12]  // argb
-    mov        ecx, [esp + 4 + 16]  // width
-    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
-    pxor       xmm4, xmm4
-
-    align      4
- convertloop:
-    READNV12
-    YVUTORGB
-
-    // Step 3: Weave into ARGB
-    punpcklbw  xmm0, xmm1           // BG
-    punpcklbw  xmm2, xmm5           // RA
-    movdqa     xmm1, xmm0
-    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
-    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
-    movdqa     [edx], xmm0
-    movdqa     [edx + 16], xmm1
-    lea        edx,  [edx + 32]
-    sub        ecx, 8
-    jg         convertloop
-
-    pop        esi
-    ret
-  }
-}
-
-// 8 pixels, unaligned.
-// 8 UV values, mixed with 8 Y producing 8 ARGB (32 bytes).
-__declspec(naked) __declspec(align(16))
-void I444ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
-                                   const uint8* u_buf,
-                                   const uint8* v_buf,
-                                   uint8* dst_argb,
-                                   int width) {
-  __asm {
-    push       esi
-    push       edi
-    mov        eax, [esp + 8 + 4]   // Y
-    mov        esi, [esp + 8 + 8]   // U
-    mov        edi, [esp + 8 + 12]  // V
-    mov        edx, [esp + 8 + 16]  // argb
-    mov        ecx, [esp + 8 + 20]  // width
-    sub        edi, esi
-    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
-    pxor       xmm4, xmm4
-
-    align      4
- convertloop:
-    READYUV444
-    YUVTORGB
-
-    // Step 3: Weave into ARGB
-    punpcklbw  xmm0, xmm1           // BG
-    punpcklbw  xmm2, xmm5           // RA
-    movdqa     xmm1, xmm0
-    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
-    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
-    movdqu     [edx], xmm0
-    movdqu     [edx + 16], xmm1
-    lea        edx,  [edx + 32]
-    sub        ecx, 8
-    jg         convertloop
-
-    pop        edi
-    pop        esi
-    ret
-  }
-}
-
-// 8 pixels, unaligned.
-// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
-__declspec(naked) __declspec(align(16))
-void I422ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
-                                   const uint8* u_buf,
-                                   const uint8* v_buf,
-                                   uint8* dst_argb,
-                                   int width) {
-  __asm {
-    push       esi
-    push       edi
-    mov        eax, [esp + 8 + 4]   // Y
-    mov        esi, [esp + 8 + 8]   // U
-    mov        edi, [esp + 8 + 12]  // V
-    mov        edx, [esp + 8 + 16]  // argb
-    mov        ecx, [esp + 8 + 20]  // width
-    sub        edi, esi
-    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
-    pxor       xmm4, xmm4
-
-    align      4
- convertloop:
-    READYUV422
-    YUVTORGB
-
-    // Step 3: Weave into ARGB
-    punpcklbw  xmm0, xmm1           // BG
-    punpcklbw  xmm2, xmm5           // RA
-    movdqa     xmm1, xmm0
-    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
-    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
-    movdqu     [edx], xmm0
-    movdqu     [edx + 16], xmm1
-    lea        edx,  [edx + 32]
-    sub        ecx, 8
-    jg         convertloop
-
-    pop        edi
-    pop        esi
-    ret
-  }
-}
-
-// 8 pixels, unaligned.
-// 2 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
-// Similar to I420 but duplicate UV once more.
-__declspec(naked) __declspec(align(16))
-void I411ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
-                                   const uint8* u_buf,
-                                   const uint8* v_buf,
-                                   uint8* dst_argb,
-                                   int width) {
-  __asm {
-    push       ebx
-    push       esi
-    push       edi
-    mov        eax, [esp + 12 + 4]   // Y
-    mov        esi, [esp + 12 + 8]   // U
-    mov        edi, [esp + 12 + 12]  // V
-    mov        edx, [esp + 12 + 16]  // argb
-    mov        ecx, [esp + 12 + 20]  // width
-    sub        edi, esi
-    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
-    pxor       xmm4, xmm4
-
-    align      4
- convertloop:
-    READYUV411  // modifies EBX
-    YUVTORGB
-
-    // Step 3: Weave into ARGB
-    punpcklbw  xmm0, xmm1           // BG
-    punpcklbw  xmm2, xmm5           // RA
-    movdqa     xmm1, xmm0
-    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
-    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
-    movdqu     [edx], xmm0
-    movdqu     [edx + 16], xmm1
-    lea        edx,  [edx + 32]
-    sub        ecx, 8
-    jg         convertloop
-
-    pop        edi
-    pop        esi
-    pop        ebx
-    ret
-  }
-}
-
-// 8 pixels, dest aligned 16.
-// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
-__declspec(naked) __declspec(align(16))
-void NV12ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
-                                   const uint8* uv_buf,
-                                   uint8* dst_argb,
-                                   int width) {
-  __asm {
-    push       esi
-    mov        eax, [esp + 4 + 4]   // Y
     mov        esi, [esp + 4 + 8]   // UV
     mov        edx, [esp + 4 + 12]  // argb
     mov        ecx, [esp + 4 + 16]  // width
     pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
-    pxor       xmm4, xmm4
 
-    align      4
  convertloop:
     READNV12
-    YUVTORGB
+    YUVTORGB(kYvuConstants)
+    STOREARGB
 
-    // Step 3: Weave into ARGB
-    punpcklbw  xmm0, xmm1           // BG
-    punpcklbw  xmm2, xmm5           // RA
-    movdqa     xmm1, xmm0
-    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
-    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
-    movdqu     [edx], xmm0
-    movdqu     [edx + 16], xmm1
-    lea        edx,  [edx + 32]
-    sub        ecx, 8
-    jg         convertloop
-
-    pop        esi
-    ret
-  }
-}
-
-// 8 pixels, dest aligned 16.
-// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
-__declspec(naked) __declspec(align(16))
-void NV21ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
-                                   const uint8* uv_buf,
-                                   uint8* dst_argb,
-                                   int width) {
-  __asm {
-    push       esi
-    mov        eax, [esp + 4 + 4]   // Y
-    mov        esi, [esp + 4 + 8]   // VU
-    mov        edx, [esp + 4 + 12]  // argb
-    mov        ecx, [esp + 4 + 16]  // width
-    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
-    pxor       xmm4, xmm4
-
-    align      4
- convertloop:
-    READNV12
-    YVUTORGB
-
-    // Step 3: Weave into ARGB
-    punpcklbw  xmm0, xmm1           // BG
-    punpcklbw  xmm2, xmm5           // RA
-    movdqa     xmm1, xmm0
-    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
-    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
-    movdqu     [edx], xmm0
-    movdqu     [edx + 16], xmm1
-    lea        edx,  [edx + 32]
     sub        ecx, 8
     jg         convertloop
 
@@ -2990,64 +2355,12 @@ void I422ToBGRARow_SSSE3(const uint8* y_buf,
     mov        edx, [esp + 8 + 16]  // bgra
     mov        ecx, [esp + 8 + 20]  // width
     sub        edi, esi
-    pxor       xmm4, xmm4
-
-    align      4
- convertloop:
-    READYUV422
-    YUVTORGB
-
-    // Step 3: Weave into BGRA
-    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
-    punpcklbw  xmm1, xmm0           // GB
-    punpcklbw  xmm5, xmm2           // AR
-    movdqa     xmm0, xmm5
-    punpcklwd  xmm5, xmm1           // BGRA first 4 pixels
-    punpckhwd  xmm0, xmm1           // BGRA next 4 pixels
-    movdqa     [edx], xmm5
-    movdqa     [edx + 16], xmm0
-    lea        edx,  [edx + 32]
-    sub        ecx, 8
-    jg         convertloop
-
-    pop        edi
-    pop        esi
-    ret
-  }
-}
-
-__declspec(naked) __declspec(align(16))
-void I422ToBGRARow_Unaligned_SSSE3(const uint8* y_buf,
-                                   const uint8* u_buf,
-                                   const uint8* v_buf,
-                                   uint8* dst_bgra,
-                                   int width) {
-  __asm {
-    push       esi
-    push       edi
-    mov        eax, [esp + 8 + 4]   // Y
-    mov        esi, [esp + 8 + 8]   // U
-    mov        edi, [esp + 8 + 12]  // V
-    mov        edx, [esp + 8 + 16]  // bgra
-    mov        ecx, [esp + 8 + 20]  // width
-    sub        edi, esi
-    pxor       xmm4, xmm4
 
-    align      4
  convertloop:
     READYUV422
-    YUVTORGB
+    YUVTORGB(kYuvConstants)
+    STOREBGRA
 
-    // Step 3: Weave into BGRA
-    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
-    punpcklbw  xmm1, xmm0           // GB
-    punpcklbw  xmm5, xmm2           // AR
-    movdqa     xmm0, xmm5
-    punpcklwd  xmm5, xmm1           // BGRA first 4 pixels
-    punpckhwd  xmm0, xmm1           // BGRA next 4 pixels
-    movdqu     [edx], xmm5
-    movdqu     [edx + 16], xmm0
-    lea        edx,  [edx + 32]
     sub        ecx, 8
     jg         convertloop
 
@@ -3073,63 +2386,12 @@ void I422ToABGRRow_SSSE3(const uint8* y_buf,
     mov        ecx, [esp + 8 + 20]  // width
     sub        edi, esi
     pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
-    pxor       xmm4, xmm4
 
-    align      4
  convertloop:
     READYUV422
-    YUVTORGB
+    YUVTORGB(kYuvConstants)
+    STOREABGR
 
-    // Step 3: Weave into ARGB
-    punpcklbw  xmm2, xmm1           // RG
-    punpcklbw  xmm0, xmm5           // BA
-    movdqa     xmm1, xmm2
-    punpcklwd  xmm2, xmm0           // RGBA first 4 pixels
-    punpckhwd  xmm1, xmm0           // RGBA next 4 pixels
-    movdqa     [edx], xmm2
-    movdqa     [edx + 16], xmm1
-    lea        edx,  [edx + 32]
-    sub        ecx, 8
-    jg         convertloop
-
-    pop        edi
-    pop        esi
-    ret
-  }
-}
-
-__declspec(naked) __declspec(align(16))
-void I422ToABGRRow_Unaligned_SSSE3(const uint8* y_buf,
-                                   const uint8* u_buf,
-                                   const uint8* v_buf,
-                                   uint8* dst_abgr,
-                                   int width) {
-  __asm {
-    push       esi
-    push       edi
-    mov        eax, [esp + 8 + 4]   // Y
-    mov        esi, [esp + 8 + 8]   // U
-    mov        edi, [esp + 8 + 12]  // V
-    mov        edx, [esp + 8 + 16]  // abgr
-    mov        ecx, [esp + 8 + 20]  // width
-    sub        edi, esi
-    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
-    pxor       xmm4, xmm4
-
-    align      4
- convertloop:
-    READYUV422
-    YUVTORGB
-
-    // Step 3: Weave into ARGB
-    punpcklbw  xmm2, xmm1           // RG
-    punpcklbw  xmm0, xmm5           // BA
-    movdqa     xmm1, xmm2
-    punpcklwd  xmm2, xmm0           // RGBA first 4 pixels
-    punpckhwd  xmm1, xmm0           // RGBA next 4 pixels
-    movdqu     [edx], xmm2
-    movdqu     [edx + 16], xmm1
-    lea        edx,  [edx + 32]
     sub        ecx, 8
     jg         convertloop
 
@@ -3154,64 +2416,12 @@ void I422ToRGBARow_SSSE3(const uint8* y_buf,
     mov        edx, [esp + 8 + 16]  // rgba
     mov        ecx, [esp + 8 + 20]  // width
     sub        edi, esi
-    pxor       xmm4, xmm4
 
-    align      4
  convertloop:
     READYUV422
-    YUVTORGB
-
-    // Step 3: Weave into RGBA
-    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
-    punpcklbw  xmm1, xmm2           // GR
-    punpcklbw  xmm5, xmm0           // AB
-    movdqa     xmm0, xmm5
-    punpcklwd  xmm5, xmm1           // RGBA first 4 pixels
-    punpckhwd  xmm0, xmm1           // RGBA next 4 pixels
-    movdqa     [edx], xmm5
-    movdqa     [edx + 16], xmm0
-    lea        edx,  [edx + 32]
-    sub        ecx, 8
-    jg         convertloop
-
-    pop        edi
-    pop        esi
-    ret
-  }
-}
+    YUVTORGB(kYuvConstants)
+    STORERGBA
 
-__declspec(naked) __declspec(align(16))
-void I422ToRGBARow_Unaligned_SSSE3(const uint8* y_buf,
-                                   const uint8* u_buf,
-                                   const uint8* v_buf,
-                                   uint8* dst_rgba,
-                                   int width) {
-  __asm {
-    push       esi
-    push       edi
-    mov        eax, [esp + 8 + 4]   // Y
-    mov        esi, [esp + 8 + 8]   // U
-    mov        edi, [esp + 8 + 12]  // V
-    mov        edx, [esp + 8 + 16]  // rgba
-    mov        ecx, [esp + 8 + 20]  // width
-    sub        edi, esi
-    pxor       xmm4, xmm4
-
-    align      4
- convertloop:
-    READYUV422
-    YUVTORGB
-
-    // Step 3: Weave into RGBA
-    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
-    punpcklbw  xmm1, xmm2           // GR
-    punpcklbw  xmm5, xmm0           // AB
-    movdqa     xmm0, xmm5
-    punpcklwd  xmm5, xmm1           // RGBA first 4 pixels
-    punpckhwd  xmm0, xmm1           // RGBA next 4 pixels
-    movdqu     [edx], xmm5
-    movdqu     [edx + 16], xmm0
-    lea        edx,  [edx + 32]
     sub        ecx, 8
     jg         convertloop
 
@@ -3224,32 +2434,32 @@ void I422ToRGBARow_Unaligned_SSSE3(const uint8* y_buf,
 #endif  // HAS_I422TOARGBROW_SSSE3
 
 #ifdef HAS_YTOARGBROW_SSE2
+// 8 pixels of Y converted to 8 pixels of ARGB (32 bytes).
 __declspec(naked) __declspec(align(16))
 void YToARGBRow_SSE2(const uint8* y_buf,
                      uint8* rgb_buf,
                      int width) {
   __asm {
-    pxor       xmm5, xmm5
-    pcmpeqb    xmm4, xmm4           // generate mask 0xff000000
-    pslld      xmm4, 24
-    mov        eax, 0x00100010
-    movd       xmm3, eax
-    pshufd     xmm3, xmm3, 0
-    mov        eax, 0x004a004a       // 74
+    mov        eax, 0x4a354a35      // 4a35 = 18997 = round(1.164 * 64 * 256)
     movd       xmm2, eax
     pshufd     xmm2, xmm2,0
+    mov        eax, 0x04880488      // 0488 = 1160 = round(1.164 * 64 * 16)
+    movd       xmm3, eax
+    pshufd     xmm3, xmm3, 0
+    pcmpeqb    xmm4, xmm4           // generate mask 0xff000000
+    pslld      xmm4, 24
+
     mov        eax, [esp + 4]       // Y
     mov        edx, [esp + 8]       // rgb
     mov        ecx, [esp + 12]      // width
 
-    align      4
  convertloop:
     // Step 1: Scale Y contribution to 8 G values. G = (y - 16) * 1.164
     movq       xmm0, qword ptr [eax]
     lea        eax, [eax + 8]
-    punpcklbw  xmm0, xmm5           // 0.Y
+    punpcklbw  xmm0, xmm0           // Y.Y
+    pmulhuw    xmm0, xmm2
     psubusw    xmm0, xmm3
-    pmullw     xmm0, xmm2
     psrlw      xmm0, 6
     packuswb   xmm0, xmm0           // G
 
@@ -3260,23 +2470,74 @@ void YToARGBRow_SSE2(const uint8* y_buf,
     punpckhwd  xmm1, xmm1           // BGRA next 4 pixels
     por        xmm0, xmm4
     por        xmm1, xmm4
-    movdqa     [edx], xmm0
-    movdqa     [edx + 16], xmm1
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
     lea        edx,  [edx + 32]
     sub        ecx, 8
     jg         convertloop
-
     ret
   }
 }
 #endif  // HAS_YTOARGBROW_SSE2
 
+#ifdef HAS_YTOARGBROW_AVX2
+// 16 pixels of Y converted to 16 pixels of ARGB (64 bytes).
+// note: vpunpcklbw mutates and vpackuswb unmutates.
+__declspec(naked) __declspec(align(16))
+void YToARGBRow_AVX2(const uint8* y_buf,
+                     uint8* rgb_buf,
+                     int width) {
+  __asm {
+    mov        eax, 0x4a354a35      // 4a35 = 18997 = round(1.164 * 64 * 256)
+    vmovd      xmm2, eax
+    vbroadcastss ymm2, xmm2
+    mov        eax, 0x04880488      // 0488 = 1160 = round(1.164 * 64 * 16)
+    vmovd      xmm3, eax
+    vbroadcastss ymm3, xmm3
+    vpcmpeqb   ymm4, ymm4, ymm4     // generate mask 0xff000000
+    vpslld     ymm4, ymm4, 24
+
+    mov        eax, [esp + 4]       // Y
+    mov        edx, [esp + 8]       // rgb
+    mov        ecx, [esp + 12]      // width
+
+ convertloop:
+    // Step 1: Scale Y contribution to 16 G values. G = (y - 16) * 1.164
+    vmovdqu    xmm0, [eax]
+    lea        eax, [eax + 16]
+    vpermq     ymm0, ymm0, 0xd8           // vpunpcklbw mutates
+    vpunpcklbw ymm0, ymm0, ymm0           // Y.Y
+    vpmulhuw   ymm0, ymm0, ymm2
+    vpsubusw   ymm0, ymm0, ymm3
+    vpsrlw     ymm0, ymm0, 6
+    vpackuswb  ymm0, ymm0, ymm0           // G.  still mutated: 3120
+
+    // TODO(fbarchard): Weave alpha with unpack.
+    // Step 2: Weave into ARGB
+    vpunpcklbw ymm1, ymm0, ymm0           // GG - mutates
+    vpermq     ymm1, ymm1, 0xd8
+    vpunpcklwd ymm0, ymm1, ymm1           // GGGG first 8 pixels
+    vpunpckhwd ymm1, ymm1, ymm1           // GGGG next 8 pixels
+    vpor       ymm0, ymm0, ymm4
+    vpor       ymm1, ymm1, ymm4
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + 32], ymm1
+    lea        edx,  [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_YTOARGBROW_AVX2
+
 #ifdef HAS_MIRRORROW_SSSE3
 // Shuffle table for reversing the bytes.
 static const uvec8 kShuffleMirror = {
   15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
 };
 
+// TODO(fbarchard): Replace lea with -16 offset.
 __declspec(naked) __declspec(align(16))
 void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
   __asm {
@@ -3284,15 +2545,13 @@ void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
     mov       edx, [esp + 8]   // dst
     mov       ecx, [esp + 12]  // width
     movdqa    xmm5, kShuffleMirror
-    lea       eax, [eax - 16]
 
-    align      4
  convertloop:
-    movdqa    xmm0, [eax + ecx]
+    movdqu    xmm0, [eax - 16 + ecx]
     pshufb    xmm0, xmm5
-    sub       ecx, 16
-    movdqa    [edx], xmm0
+    movdqu    [edx], xmm0
     lea       edx, [edx + 16]
+    sub       ecx, 16
     jg        convertloop
     ret
   }
@@ -3300,29 +2559,21 @@ void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
 #endif  // HAS_MIRRORROW_SSSE3
 
 #ifdef HAS_MIRRORROW_AVX2
-// Shuffle table for reversing the bytes.
-static const ulvec8 kShuffleMirror_AVX2 = {
-  15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u,
-  15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
-};
-
 __declspec(naked) __declspec(align(16))
 void MirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
   __asm {
     mov       eax, [esp + 4]   // src
     mov       edx, [esp + 8]   // dst
     mov       ecx, [esp + 12]  // width
-    vmovdqa   ymm5, kShuffleMirror_AVX2
-    lea       eax, [eax - 32]
+    vbroadcastf128 ymm5, kShuffleMirror
 
-    align      4
  convertloop:
-    vmovdqu   ymm0, [eax + ecx]
+    vmovdqu   ymm0, [eax - 32 + ecx]
     vpshufb   ymm0, ymm0, ymm5
     vpermq    ymm0, ymm0, 0x4e  // swap high and low halfs
-    sub       ecx, 32
     vmovdqu   [edx], ymm0
     lea       edx, [edx + 32]
+    sub       ecx, 32
     jg        convertloop
     vzeroupper
     ret
@@ -3331,19 +2582,15 @@ void MirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
 #endif  // HAS_MIRRORROW_AVX2
 
 #ifdef HAS_MIRRORROW_SSE2
-// SSE2 version has movdqu so it can be used on unaligned buffers when SSSE3
-// version can not.
 __declspec(naked) __declspec(align(16))
 void MirrorRow_SSE2(const uint8* src, uint8* dst, int width) {
   __asm {
     mov       eax, [esp + 4]   // src
     mov       edx, [esp + 8]   // dst
     mov       ecx, [esp + 12]  // width
-    lea       eax, [eax - 16]
 
-    align      4
  convertloop:
-    movdqu    xmm0, [eax + ecx]
+    movdqu    xmm0, [eax - 16 + ecx]
     movdqa    xmm1, xmm0        // swap bytes
     psllw     xmm0, 8
     psrlw     xmm1, 8
@@ -3351,9 +2598,9 @@ void MirrorRow_SSE2(const uint8* src, uint8* dst, int width) {
     pshuflw   xmm0, xmm0, 0x1b  // swap words
     pshufhw   xmm0, xmm0, 0x1b
     pshufd    xmm0, xmm0, 0x4e  // swap qwords
-    sub       ecx, 16
     movdqu    [edx], xmm0
     lea       edx, [edx + 16]
+    sub       ecx, 16
     jg        convertloop
     ret
   }
@@ -3379,15 +2626,14 @@ void MirrorUVRow_SSSE3(const uint8* src, uint8* dst_u, uint8* dst_v,
     lea       eax, [eax + ecx * 2 - 16]
     sub       edi, edx
 
-    align      4
  convertloop:
-    movdqa    xmm0, [eax]
+    movdqu    xmm0, [eax]
     lea       eax, [eax - 16]
     pshufb    xmm0, xmm1
-    sub       ecx, 8
     movlpd    qword ptr [edx], xmm0
     movhpd    qword ptr [edx + edi], xmm0
     lea       edx, [edx + 8]
+    sub       ecx, 8
     jg        convertloop
 
     pop       edi
@@ -3396,34 +2642,27 @@ void MirrorUVRow_SSSE3(const uint8* src, uint8* dst_u, uint8* dst_v,
 }
 #endif  // HAS_MIRRORROW_UV_SSSE3
 
-#ifdef HAS_ARGBMIRRORROW_SSSE3
-// Shuffle table for reversing the bytes.
-static const uvec8 kARGBShuffleMirror = {
-  12u, 13u, 14u, 15u, 8u, 9u, 10u, 11u, 4u, 5u, 6u, 7u, 0u, 1u, 2u, 3u
-};
-
+#ifdef HAS_ARGBMIRRORROW_SSE2
 __declspec(naked) __declspec(align(16))
-void ARGBMirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
+void ARGBMirrorRow_SSE2(const uint8* src, uint8* dst, int width) {
   __asm {
     mov       eax, [esp + 4]   // src
     mov       edx, [esp + 8]   // dst
     mov       ecx, [esp + 12]  // width
     lea       eax, [eax - 16 + ecx * 4]  // last 4 pixels.
-    movdqa    xmm5, kARGBShuffleMirror
 
-    align      4
  convertloop:
-    movdqa    xmm0, [eax]
+    movdqu    xmm0, [eax]
     lea       eax, [eax - 16]
-    pshufb    xmm0, xmm5
-    sub       ecx, 4
-    movdqa    [edx], xmm0
+    pshufd    xmm0, xmm0, 0x1b
+    movdqu    [edx], xmm0
     lea       edx, [edx + 16]
+    sub       ecx, 4
     jg        convertloop
     ret
   }
 }
-#endif  // HAS_ARGBMIRRORROW_SSSE3
+#endif  // HAS_ARGBMIRRORROW_SSE2
 
 #ifdef HAS_ARGBMIRRORROW_AVX2
 // Shuffle table for reversing the bytes.
@@ -3437,15 +2676,13 @@ void ARGBMirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
     mov       eax, [esp + 4]   // src
     mov       edx, [esp + 8]   // dst
     mov       ecx, [esp + 12]  // width
-    lea       eax, [eax - 32]
-    vmovdqa   ymm5, kARGBShuffleMirror_AVX2
+    vmovdqu   ymm5, kARGBShuffleMirror_AVX2
 
-    align      4
  convertloop:
-    vpermd    ymm0, ymm5, [eax + ecx * 4]  // permute dword order
-    sub       ecx, 8
+    vpermd    ymm0, ymm5, [eax - 32 + ecx * 4]  // permute dword order
     vmovdqu   [edx], ymm0
     lea       edx, [edx + 32]
+    sub       ecx, 8
     jg        convertloop
     vzeroupper
     ret
@@ -3466,44 +2703,6 @@ void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) {
     psrlw      xmm5, 8
     sub        edi, edx
 
-    align      4
-  convertloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    lea        eax,  [eax + 32]
-    movdqa     xmm2, xmm0
-    movdqa     xmm3, xmm1
-    pand       xmm0, xmm5   // even bytes
-    pand       xmm1, xmm5
-    packuswb   xmm0, xmm1
-    psrlw      xmm2, 8      // odd bytes
-    psrlw      xmm3, 8
-    packuswb   xmm2, xmm3
-    movdqa     [edx], xmm0
-    movdqa     [edx + edi], xmm2
-    lea        edx, [edx + 16]
-    sub        ecx, 16
-    jg         convertloop
-
-    pop        edi
-    ret
-  }
-}
-
-__declspec(naked) __declspec(align(16))
-void SplitUVRow_Unaligned_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
-                               int pix) {
-  __asm {
-    push       edi
-    mov        eax, [esp + 4 + 4]    // src_uv
-    mov        edx, [esp + 4 + 8]    // dst_u
-    mov        edi, [esp + 4 + 12]   // dst_v
-    mov        ecx, [esp + 4 + 16]   // pix
-    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
-    psrlw      xmm5, 8
-    sub        edi, edx
-
-    align      4
   convertloop:
     movdqu     xmm0, [eax]
     movdqu     xmm1, [eax + 16]
@@ -3526,6 +2725,7 @@ void SplitUVRow_Unaligned_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
     ret
   }
 }
+
 #endif  // HAS_SPLITUVROW_SSE2
 
 #ifdef HAS_SPLITUVROW_AVX2
@@ -3541,7 +2741,6 @@ void SplitUVRow_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) {
     vpsrlw     ymm5, ymm5, 8
     sub        edi, edx
 
-    align      4
   convertloop:
     vmovdqu    ymm0, [eax]
     vmovdqu    ymm1, [eax + 32]
@@ -3579,37 +2778,6 @@ void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
     mov        ecx, [esp + 4 + 16]   // width
     sub        edx, eax
 
-    align      4
-  convertloop:
-    movdqa     xmm0, [eax]      // read 16 U's
-    movdqa     xmm1, [eax + edx]  // and 16 V's
-    lea        eax,  [eax + 16]
-    movdqa     xmm2, xmm0
-    punpcklbw  xmm0, xmm1       // first 8 UV pairs
-    punpckhbw  xmm2, xmm1       // next 8 UV pairs
-    movdqa     [edi], xmm0
-    movdqa     [edi + 16], xmm2
-    lea        edi, [edi + 32]
-    sub        ecx, 16
-    jg         convertloop
-
-    pop        edi
-    ret
-  }
-}
-
-__declspec(naked) __declspec(align(16))
-void MergeUVRow_Unaligned_SSE2(const uint8* src_u, const uint8* src_v,
-                               uint8* dst_uv, int width) {
-  __asm {
-    push       edi
-    mov        eax, [esp + 4 + 4]    // src_u
-    mov        edx, [esp + 4 + 8]    // src_v
-    mov        edi, [esp + 4 + 12]   // dst_uv
-    mov        ecx, [esp + 4 + 16]   // width
-    sub        edx, eax
-
-    align      4
   convertloop:
     movdqu     xmm0, [eax]      // read 16 U's
     movdqu     xmm1, [eax + edx]  // and 16 V's
@@ -3641,17 +2809,16 @@ void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
     mov        ecx, [esp + 4 + 16]   // width
     sub        edx, eax
 
-    align      4
   convertloop:
     vmovdqu    ymm0, [eax]           // read 32 U's
     vmovdqu    ymm1, [eax + edx]     // and 32 V's
     lea        eax,  [eax + 32]
     vpunpcklbw ymm2, ymm0, ymm1      // low 16 UV pairs. mutated qqword 0,2
     vpunpckhbw ymm0, ymm0, ymm1      // high 16 UV pairs. mutated qqword 1,3
-    vperm2i128 ymm1, ymm2, ymm0, 0x20  // low 128 of ymm2 and low 128 of ymm0
-    vperm2i128 ymm2, ymm2, ymm0, 0x31  // high 128 of ymm2 and high 128 of ymm0
-    vmovdqu    [edi], ymm1
-    vmovdqu    [edi + 32], ymm2
+    vextractf128 [edi], ymm2, 0       // bytes 0..15
+    vextractf128 [edi + 16], ymm0, 0  // bytes 16..31
+    vextractf128 [edi + 32], ymm2, 1  // bytes 32..47
+    vextractf128 [edi + 48], ymm0, 1  // bytes 47..63
     lea        edi, [edi + 64]
     sub        ecx, 32
     jg         convertloop
@@ -3672,13 +2839,12 @@ void CopyRow_SSE2(const uint8* src, uint8* dst, int count) {
     mov        edx, [esp + 8]   // dst
     mov        ecx, [esp + 12]  // count
 
-    align      4
   convertloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
     lea        eax, [eax + 32]
-    movdqa     [edx], xmm0
-    movdqa     [edx + 16], xmm1
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
     lea        edx, [edx + 32]
     sub        ecx, 32
     jg         convertloop
@@ -3687,39 +2853,46 @@ void CopyRow_SSE2(const uint8* src, uint8* dst, int count) {
 }
 #endif  // HAS_COPYROW_SSE2
 
-// Unaligned Multiple of 1.
+#ifdef HAS_COPYROW_AVX
+// CopyRow copys 'count' bytes using a 32 byte load/store, 64 bytes at time.
 __declspec(naked) __declspec(align(16))
-void CopyRow_ERMS(const uint8* src, uint8* dst, int count) {
+void CopyRow_AVX(const uint8* src, uint8* dst, int count) {
   __asm {
-    mov        eax, esi
-    mov        edx, edi
-    mov        esi, [esp + 4]   // src
-    mov        edi, [esp + 8]   // dst
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
     mov        ecx, [esp + 12]  // count
-    rep movsb
-    mov        edi, edx
-    mov        esi, eax
+
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax, [eax + 64]
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + 32], ymm1
+    lea        edx, [edx + 64]
+    sub        ecx, 64
+    jg         convertloop
+
+    vzeroupper
     ret
   }
 }
+#endif  // HAS_COPYROW_AVX
 
-#ifdef HAS_COPYROW_X86
+// Multiple of 1.
 __declspec(naked) __declspec(align(16))
-void CopyRow_X86(const uint8* src, uint8* dst, int count) {
+void CopyRow_ERMS(const uint8* src, uint8* dst, int count) {
   __asm {
     mov        eax, esi
     mov        edx, edi
     mov        esi, [esp + 4]   // src
     mov        edi, [esp + 8]   // dst
     mov        ecx, [esp + 12]  // count
-    shr        ecx, 2
-    rep movsd
+    rep movsb
     mov        edi, edx
     mov        esi, eax
     ret
   }
 }
-#endif  // HAS_COPYROW_X86
 
 #ifdef HAS_ARGBCOPYALPHAROW_SSE2
 // width in pixels
@@ -3734,21 +2907,20 @@ void ARGBCopyAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
     pcmpeqb    xmm1, xmm1       // generate mask 0x00ffffff
     psrld      xmm1, 8
 
-    align      4
   convertloop:
-    movdqa     xmm2, [eax]
-    movdqa     xmm3, [eax + 16]
+    movdqu     xmm2, [eax]
+    movdqu     xmm3, [eax + 16]
     lea        eax, [eax + 32]
-    movdqa     xmm4, [edx]
-    movdqa     xmm5, [edx + 16]
+    movdqu     xmm4, [edx]
+    movdqu     xmm5, [edx + 16]
     pand       xmm2, xmm0
     pand       xmm3, xmm0
     pand       xmm4, xmm1
     pand       xmm5, xmm1
     por        xmm2, xmm4
     por        xmm3, xmm5
-    movdqa     [edx], xmm2
-    movdqa     [edx + 16], xmm3
+    movdqu     [edx], xmm2
+    movdqu     [edx + 16], xmm3
     lea        edx, [edx + 32]
     sub        ecx, 8
     jg         convertloop
@@ -3769,7 +2941,6 @@ void ARGBCopyAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
     vpcmpeqb   ymm0, ymm0, ymm0
     vpsrld     ymm0, ymm0, 8    // generate mask 0x00ffffff
 
-    align      4
   convertloop:
     vmovdqu    ymm1, [eax]
     vmovdqu    ymm2, [eax + 32]
@@ -3801,23 +2972,22 @@ void ARGBCopyYToAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
     pcmpeqb    xmm1, xmm1       // generate mask 0x00ffffff
     psrld      xmm1, 8
 
-    align      4
   convertloop:
     movq       xmm2, qword ptr [eax]  // 8 Y's
     lea        eax, [eax + 8]
     punpcklbw  xmm2, xmm2
     punpckhwd  xmm3, xmm2
     punpcklwd  xmm2, xmm2
-    movdqa     xmm4, [edx]
-    movdqa     xmm5, [edx + 16]
+    movdqu     xmm4, [edx]
+    movdqu     xmm5, [edx + 16]
     pand       xmm2, xmm0
     pand       xmm3, xmm0
     pand       xmm4, xmm1
     pand       xmm5, xmm1
     por        xmm2, xmm4
     por        xmm3, xmm5
-    movdqa     [edx], xmm2
-    movdqa     [edx + 16], xmm3
+    movdqu     [edx], xmm2
+    movdqu     [edx + 16], xmm3
     lea        edx, [edx + 32]
     sub        ecx, 8
     jg         convertloop
@@ -3838,7 +3008,6 @@ void ARGBCopyYToAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
     vpcmpeqb   ymm0, ymm0, ymm0
     vpsrld     ymm0, ymm0, 8    // generate mask 0x00ffffff
 
-    align      4
   convertloop:
     vpmovzxbd  ymm1, qword ptr [eax]
     vpmovzxbd  ymm2, qword ptr [eax + 8]
@@ -3860,13 +3029,16 @@ void ARGBCopyYToAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
 #endif  // HAS_ARGBCOPYYTOALPHAROW_AVX2
 
 #ifdef HAS_SETROW_X86
-// SetRow8 writes 'count' bytes using a 32 bit value repeated.
+// Write 'count' bytes using an 8 bit value repeated.
+// Count should be multiple of 4.
 __declspec(naked) __declspec(align(16))
-void SetRow_X86(uint8* dst, uint32 v32, int count) {
+void SetRow_X86(uint8* dst, uint8 v8, int count) {
   __asm {
+    movzx      eax, byte ptr [esp + 8]    // v8
+    mov        edx, 0x01010101  // Duplicate byte to all bytes.
+    mul        edx              // overwrites edx with upper part of result.
     mov        edx, edi
     mov        edi, [esp + 4]   // dst
-    mov        eax, [esp + 8]   // v32
     mov        ecx, [esp + 12]  // count
     shr        ecx, 2
     rep stosd
@@ -3875,33 +3047,30 @@ void SetRow_X86(uint8* dst, uint32 v32, int count) {
   }
 }
 
-// SetRow32 writes 'count' words using a 32 bit value repeated.
+// Write 'count' bytes using an 8 bit value repeated.
 __declspec(naked) __declspec(align(16))
-void ARGBSetRows_X86(uint8* dst, uint32 v32, int width,
-                   int dst_stride, int height) {
+void SetRow_ERMS(uint8* dst, uint8 v8, int count) {
   __asm {
-    push       esi
-    push       edi
-    push       ebp
-    mov        edi, [esp + 12 + 4]   // dst
-    mov        eax, [esp + 12 + 8]   // v32
-    mov        ebp, [esp + 12 + 12]  // width
-    mov        edx, [esp + 12 + 16]  // dst_stride
-    mov        esi, [esp + 12 + 20]  // height
-    lea        ecx, [ebp * 4]
-    sub        edx, ecx             // stride - width * 4
-
-    align      4
-  convertloop:
-    mov        ecx, ebp
-    rep stosd
-    add        edi, edx
-    sub        esi, 1
-    jg         convertloop
+    mov        edx, edi
+    mov        edi, [esp + 4]   // dst
+    mov        eax, [esp + 8]   // v8
+    mov        ecx, [esp + 12]  // count
+    rep stosb
+    mov        edi, edx
+    ret
+  }
+}
 
-    pop        ebp
-    pop        edi
-    pop        esi
+// Write 'count' 32 bit values.
+__declspec(naked) __declspec(align(16))
+void ARGBSetRow_X86(uint8* dst_argb, uint32 v32, int count) {
+  __asm {
+    mov        edx, edi
+    mov        edi, [esp + 4]   // dst
+    mov        eax, [esp + 8]   // v32
+    mov        ecx, [esp + 12]  // count
+    rep stosd
+    mov        edi, edx
     ret
   }
 }
@@ -3918,7 +3087,6 @@ void YUY2ToYRow_AVX2(const uint8* src_yuy2,
     vpcmpeqb   ymm5, ymm5, ymm5  // generate mask 0x00ff00ff
     vpsrlw     ymm5, ymm5, 8
 
-    align      4
   convertloop:
     vmovdqu    ymm0, [eax]
     vmovdqu    ymm1, [eax + 32]
@@ -3927,9 +3095,9 @@ void YUY2ToYRow_AVX2(const uint8* src_yuy2,
     vpand      ymm1, ymm1, ymm5
     vpackuswb  ymm0, ymm0, ymm1   // mutates.
     vpermq     ymm0, ymm0, 0xd8
-    sub        ecx, 32
     vmovdqu    [edx], ymm0
     lea        edx, [edx + 32]
+    sub        ecx, 32
     jg         convertloop
     vzeroupper
     ret
@@ -3951,7 +3119,6 @@ void YUY2ToUVRow_AVX2(const uint8* src_yuy2, int stride_yuy2,
     vpsrlw     ymm5, ymm5, 8
     sub        edi, edx
 
-    align      4
   convertloop:
     vmovdqu    ymm0, [eax]
     vmovdqu    ymm1, [eax + 32]
@@ -3994,7 +3161,6 @@ void YUY2ToUV422Row_AVX2(const uint8* src_yuy2,
     vpsrlw     ymm5, ymm5, 8
     sub        edi, edx
 
-    align      4
   convertloop:
     vmovdqu    ymm0, [eax]
     vmovdqu    ymm1, [eax + 32]
@@ -4029,7 +3195,6 @@ void UYVYToYRow_AVX2(const uint8* src_uyvy,
     mov        edx, [esp + 8]    // dst_y
     mov        ecx, [esp + 12]   // pix
 
-    align      4
   convertloop:
     vmovdqu    ymm0, [eax]
     vmovdqu    ymm1, [eax + 32]
@@ -4038,12 +3203,12 @@ void UYVYToYRow_AVX2(const uint8* src_uyvy,
     vpsrlw     ymm1, ymm1, 8
     vpackuswb  ymm0, ymm0, ymm1   // mutates.
     vpermq     ymm0, ymm0, 0xd8
-    sub        ecx, 32
     vmovdqu    [edx], ymm0
     lea        edx, [edx + 32]
+    sub        ecx, 32
     jg         convertloop
-    ret
     vzeroupper
+    ret
   }
 }
 
@@ -4062,7 +3227,6 @@ void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
     vpsrlw     ymm5, ymm5, 8
     sub        edi, edx
 
-    align      4
   convertloop:
     vmovdqu    ymm0, [eax]
     vmovdqu    ymm1, [eax + 32]
@@ -4105,7 +3269,6 @@ void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
     vpsrlw     ymm5, ymm5, 8
     sub        edi, edx
 
-    align      4
   convertloop:
     vmovdqu    ymm0, [eax]
     vmovdqu    ymm1, [eax + 32]
@@ -4144,114 +3307,6 @@ void YUY2ToYRow_SSE2(const uint8* src_yuy2,
     pcmpeqb    xmm5, xmm5        // generate mask 0x00ff00ff
     psrlw      xmm5, 8
 
-    align      4
-  convertloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    lea        eax,  [eax + 32]
-    pand       xmm0, xmm5   // even bytes are Y
-    pand       xmm1, xmm5
-    packuswb   xmm0, xmm1
-    sub        ecx, 16
-    movdqa     [edx], xmm0
-    lea        edx, [edx + 16]
-    jg         convertloop
-    ret
-  }
-}
-
-__declspec(naked) __declspec(align(16))
-void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
-                      uint8* dst_u, uint8* dst_v, int pix) {
-  __asm {
-    push       esi
-    push       edi
-    mov        eax, [esp + 8 + 4]    // src_yuy2
-    mov        esi, [esp + 8 + 8]    // stride_yuy2
-    mov        edx, [esp + 8 + 12]   // dst_u
-    mov        edi, [esp + 8 + 16]   // dst_v
-    mov        ecx, [esp + 8 + 20]   // pix
-    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
-    psrlw      xmm5, 8
-    sub        edi, edx
-
-    align      4
-  convertloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    movdqa     xmm2, [eax + esi]
-    movdqa     xmm3, [eax + esi + 16]
-    lea        eax,  [eax + 32]
-    pavgb      xmm0, xmm2
-    pavgb      xmm1, xmm3
-    psrlw      xmm0, 8      // YUYV -> UVUV
-    psrlw      xmm1, 8
-    packuswb   xmm0, xmm1
-    movdqa     xmm1, xmm0
-    pand       xmm0, xmm5  // U
-    packuswb   xmm0, xmm0
-    psrlw      xmm1, 8     // V
-    packuswb   xmm1, xmm1
-    movq       qword ptr [edx], xmm0
-    movq       qword ptr [edx + edi], xmm1
-    lea        edx, [edx + 8]
-    sub        ecx, 16
-    jg         convertloop
-
-    pop        edi
-    pop        esi
-    ret
-  }
-}
-
-__declspec(naked) __declspec(align(16))
-void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
-                         uint8* dst_u, uint8* dst_v, int pix) {
-  __asm {
-    push       edi
-    mov        eax, [esp + 4 + 4]    // src_yuy2
-    mov        edx, [esp + 4 + 8]    // dst_u
-    mov        edi, [esp + 4 + 12]   // dst_v
-    mov        ecx, [esp + 4 + 16]   // pix
-    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
-    psrlw      xmm5, 8
-    sub        edi, edx
-
-    align      4
-  convertloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    lea        eax,  [eax + 32]
-    psrlw      xmm0, 8      // YUYV -> UVUV
-    psrlw      xmm1, 8
-    packuswb   xmm0, xmm1
-    movdqa     xmm1, xmm0
-    pand       xmm0, xmm5  // U
-    packuswb   xmm0, xmm0
-    psrlw      xmm1, 8     // V
-    packuswb   xmm1, xmm1
-    movq       qword ptr [edx], xmm0
-    movq       qword ptr [edx + edi], xmm1
-    lea        edx, [edx + 8]
-    sub        ecx, 16
-    jg         convertloop
-
-    pop        edi
-    ret
-  }
-}
-
-__declspec(naked) __declspec(align(16))
-void YUY2ToYRow_Unaligned_SSE2(const uint8* src_yuy2,
-                               uint8* dst_y, int pix) {
-  __asm {
-    mov        eax, [esp + 4]    // src_yuy2
-    mov        edx, [esp + 8]    // dst_y
-    mov        ecx, [esp + 12]   // pix
-    pcmpeqb    xmm5, xmm5        // generate mask 0x00ff00ff
-    psrlw      xmm5, 8
-
-    align      4
   convertloop:
     movdqu     xmm0, [eax]
     movdqu     xmm1, [eax + 16]
@@ -4259,17 +3314,17 @@ void YUY2ToYRow_Unaligned_SSE2(const uint8* src_yuy2,
     pand       xmm0, xmm5   // even bytes are Y
     pand       xmm1, xmm5
     packuswb   xmm0, xmm1
-    sub        ecx, 16
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 16
     jg         convertloop
     ret
   }
 }
 
 __declspec(naked) __declspec(align(16))
-void YUY2ToUVRow_Unaligned_SSE2(const uint8* src_yuy2, int stride_yuy2,
-                                uint8* dst_u, uint8* dst_v, int pix) {
+void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
   __asm {
     push       esi
     push       edi
@@ -4282,7 +3337,6 @@ void YUY2ToUVRow_Unaligned_SSE2(const uint8* src_yuy2, int stride_yuy2,
     psrlw      xmm5, 8
     sub        edi, edx
 
-    align      4
   convertloop:
     movdqu     xmm0, [eax]
     movdqu     xmm1, [eax + 16]
@@ -4312,8 +3366,8 @@ void YUY2ToUVRow_Unaligned_SSE2(const uint8* src_yuy2, int stride_yuy2,
 }
 
 __declspec(naked) __declspec(align(16))
-void YUY2ToUV422Row_Unaligned_SSE2(const uint8* src_yuy2,
-                                   uint8* dst_u, uint8* dst_v, int pix) {
+void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix) {
   __asm {
     push       edi
     mov        eax, [esp + 4 + 4]    // src_yuy2
@@ -4324,7 +3378,6 @@ void YUY2ToUV422Row_Unaligned_SSE2(const uint8* src_yuy2,
     psrlw      xmm5, 8
     sub        edi, edx
 
-    align      4
   convertloop:
     movdqu     xmm0, [eax]
     movdqu     xmm1, [eax + 16]
@@ -4356,112 +3409,6 @@ void UYVYToYRow_SSE2(const uint8* src_uyvy,
     mov        edx, [esp + 8]    // dst_y
     mov        ecx, [esp + 12]   // pix
 
-    align      4
-  convertloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    lea        eax,  [eax + 32]
-    psrlw      xmm0, 8    // odd bytes are Y
-    psrlw      xmm1, 8
-    packuswb   xmm0, xmm1
-    sub        ecx, 16
-    movdqa     [edx], xmm0
-    lea        edx, [edx + 16]
-    jg         convertloop
-    ret
-  }
-}
-
-__declspec(naked) __declspec(align(16))
-void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
-                      uint8* dst_u, uint8* dst_v, int pix) {
-  __asm {
-    push       esi
-    push       edi
-    mov        eax, [esp + 8 + 4]    // src_yuy2
-    mov        esi, [esp + 8 + 8]    // stride_yuy2
-    mov        edx, [esp + 8 + 12]   // dst_u
-    mov        edi, [esp + 8 + 16]   // dst_v
-    mov        ecx, [esp + 8 + 20]   // pix
-    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
-    psrlw      xmm5, 8
-    sub        edi, edx
-
-    align      4
-  convertloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    movdqa     xmm2, [eax + esi]
-    movdqa     xmm3, [eax + esi + 16]
-    lea        eax,  [eax + 32]
-    pavgb      xmm0, xmm2
-    pavgb      xmm1, xmm3
-    pand       xmm0, xmm5   // UYVY -> UVUV
-    pand       xmm1, xmm5
-    packuswb   xmm0, xmm1
-    movdqa     xmm1, xmm0
-    pand       xmm0, xmm5  // U
-    packuswb   xmm0, xmm0
-    psrlw      xmm1, 8     // V
-    packuswb   xmm1, xmm1
-    movq       qword ptr [edx], xmm0
-    movq       qword ptr [edx + edi], xmm1
-    lea        edx, [edx + 8]
-    sub        ecx, 16
-    jg         convertloop
-
-    pop        edi
-    pop        esi
-    ret
-  }
-}
-
-__declspec(naked) __declspec(align(16))
-void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
-                         uint8* dst_u, uint8* dst_v, int pix) {
-  __asm {
-    push       edi
-    mov        eax, [esp + 4 + 4]    // src_yuy2
-    mov        edx, [esp + 4 + 8]    // dst_u
-    mov        edi, [esp + 4 + 12]   // dst_v
-    mov        ecx, [esp + 4 + 16]   // pix
-    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
-    psrlw      xmm5, 8
-    sub        edi, edx
-
-    align      4
-  convertloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    lea        eax,  [eax + 32]
-    pand       xmm0, xmm5   // UYVY -> UVUV
-    pand       xmm1, xmm5
-    packuswb   xmm0, xmm1
-    movdqa     xmm1, xmm0
-    pand       xmm0, xmm5  // U
-    packuswb   xmm0, xmm0
-    psrlw      xmm1, 8     // V
-    packuswb   xmm1, xmm1
-    movq       qword ptr [edx], xmm0
-    movq       qword ptr [edx + edi], xmm1
-    lea        edx, [edx + 8]
-    sub        ecx, 16
-    jg         convertloop
-
-    pop        edi
-    ret
-  }
-}
-
-__declspec(naked) __declspec(align(16))
-void UYVYToYRow_Unaligned_SSE2(const uint8* src_uyvy,
-                               uint8* dst_y, int pix) {
-  __asm {
-    mov        eax, [esp + 4]    // src_uyvy
-    mov        edx, [esp + 8]    // dst_y
-    mov        ecx, [esp + 12]   // pix
-
-    align      4
   convertloop:
     movdqu     xmm0, [eax]
     movdqu     xmm1, [eax + 16]
@@ -4469,17 +3416,17 @@ void UYVYToYRow_Unaligned_SSE2(const uint8* src_uyvy,
     psrlw      xmm0, 8    // odd bytes are Y
     psrlw      xmm1, 8
     packuswb   xmm0, xmm1
-    sub        ecx, 16
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 16
     jg         convertloop
     ret
   }
 }
 
 __declspec(naked) __declspec(align(16))
-void UYVYToUVRow_Unaligned_SSE2(const uint8* src_uyvy, int stride_uyvy,
-                                uint8* dst_u, uint8* dst_v, int pix) {
+void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
   __asm {
     push       esi
     push       edi
@@ -4492,7 +3439,6 @@ void UYVYToUVRow_Unaligned_SSE2(const uint8* src_uyvy, int stride_uyvy,
     psrlw      xmm5, 8
     sub        edi, edx
 
-    align      4
   convertloop:
     movdqu     xmm0, [eax]
     movdqu     xmm1, [eax + 16]
@@ -4522,8 +3468,8 @@ void UYVYToUVRow_Unaligned_SSE2(const uint8* src_uyvy, int stride_uyvy,
 }
 
 __declspec(naked) __declspec(align(16))
-void UYVYToUV422Row_Unaligned_SSE2(const uint8* src_uyvy,
-                                   uint8* dst_u, uint8* dst_v, int pix) {
+void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) {
   __asm {
     push       edi
     mov        eax, [esp + 4 + 4]    // src_yuy2
@@ -4534,7 +3480,6 @@ void UYVYToUV422Row_Unaligned_SSE2(const uint8* src_uyvy,
     psrlw      xmm5, 8
     sub        edi, edx
 
-    align      4
   convertloop:
     movdqu     xmm0, [eax]
     movdqu     xmm1, [eax + 16]
@@ -4607,9 +3552,9 @@ void ARGBBlendRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
     paddusb    xmm0, xmm2       // + src argb
     pand       xmm1, xmm5       // a_g_ convert to 8 bits again
     paddusb    xmm0, xmm1       // + src argb
-    sub        ecx, 1
     movd       [edx], xmm0
     lea        edx, [edx + 4]
+    sub        ecx, 1
     jge        alignloop1
 
   alignloop1b:
@@ -4638,9 +3583,9 @@ void ARGBBlendRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
     paddusb    xmm0, xmm2       // + src argb
     pand       xmm1, xmm5       // a_g_ convert to 8 bits again
     paddusb    xmm0, xmm1       // + src argb
-    sub        ecx, 4
-    movdqa     [edx], xmm0
+    movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 4
     jge        convertloop4
 
   convertloop4b:
@@ -4669,9 +3614,9 @@ void ARGBBlendRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
     paddusb    xmm0, xmm2       // + src argb
     pand       xmm1, xmm5       // a_g_ convert to 8 bits again
     paddusb    xmm0, xmm1       // + src argb
-    sub        ecx, 1
     movd       [edx], xmm0
     lea        edx, [edx + 4]
+    sub        ecx, 1
     jge        convertloop1
 
   convertloop1b:
@@ -4739,48 +3684,17 @@ void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
     paddusb    xmm0, xmm2       // + src argb
     pand       xmm1, xmm5       // a_g_ convert to 8 bits again
     paddusb    xmm0, xmm1       // + src argb
-    sub        ecx, 1
     movd       [edx], xmm0
     lea        edx, [edx + 4]
+    sub        ecx, 1
     jge        alignloop1
 
   alignloop1b:
     add        ecx, 1 - 4
     jl         convertloop4b
 
-    test       eax, 15          // unaligned?
-    jne        convertuloop4
-    test       esi, 15          // unaligned?
-    jne        convertuloop4
-
     // 4 pixel loop.
   convertloop4:
-    movdqa     xmm3, [eax]      // src argb
-    lea        eax, [eax + 16]
-    movdqa     xmm0, xmm3       // src argb
-    pxor       xmm3, xmm4       // ~alpha
-    movdqa     xmm2, [esi]      // _r_b
-    pshufb     xmm3, kShuffleAlpha // alpha
-    pand       xmm2, xmm6       // _r_b
-    paddw      xmm3, xmm7       // 256 - alpha
-    pmullw     xmm2, xmm3       // _r_b * alpha
-    movdqa     xmm1, [esi]      // _a_g
-    lea        esi, [esi + 16]
-    psrlw      xmm1, 8          // _a_g
-    por        xmm0, xmm4       // set alpha to 255
-    pmullw     xmm1, xmm3       // _a_g * alpha
-    psrlw      xmm2, 8          // _r_b convert to 8 bits again
-    paddusb    xmm0, xmm2       // + src argb
-    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
-    paddusb    xmm0, xmm1       // + src argb
-    sub        ecx, 4
-    movdqa     [edx], xmm0
-    lea        edx, [edx + 16]
-    jge        convertloop4
-    jmp        convertloop4b
-
-    // 4 pixel unaligned loop.
-  convertuloop4:
     movdqu     xmm3, [eax]      // src argb
     lea        eax, [eax + 16]
     movdqa     xmm0, xmm3       // src argb
@@ -4799,10 +3713,10 @@ void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
     paddusb    xmm0, xmm2       // + src argb
     pand       xmm1, xmm5       // a_g_ convert to 8 bits again
     paddusb    xmm0, xmm1       // + src argb
-    sub        ecx, 4
-    movdqa     [edx], xmm0
+    movdqu     [edx], xmm0
     lea        edx, [edx + 16]
-    jge        convertuloop4
+    sub        ecx, 4
+    jge        convertloop4
 
   convertloop4b:
     add        ecx, 4 - 1
@@ -4828,9 +3742,9 @@ void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
     paddusb    xmm0, xmm2       // + src argb
     pand       xmm1, xmm5       // a_g_ convert to 8 bits again
     paddusb    xmm0, xmm1       // + src argb
-    sub        ecx, 1
     movd       [edx], xmm0
     lea        edx, [edx + 4]
+    sub        ecx, 1
     jge        convertloop1
 
   convertloop1b:
@@ -4842,7 +3756,6 @@ void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
 
 #ifdef HAS_ARGBATTENUATEROW_SSE2
 // Attenuate 4 pixels at a time.
-// Aligned to 16 bytes.
 __declspec(naked) __declspec(align(16))
 void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width) {
   __asm {
@@ -4854,19 +3767,18 @@ void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width) {
     pcmpeqb    xmm5, xmm5       // generate mask 0x00ffffff
     psrld      xmm5, 8
 
-    align      4
  convertloop:
-    movdqa     xmm0, [eax]      // read 4 pixels
+    movdqu     xmm0, [eax]      // read 4 pixels
     punpcklbw  xmm0, xmm0       // first 2
     pshufhw    xmm2, xmm0, 0FFh // 8 alpha words
     pshuflw    xmm2, xmm2, 0FFh
     pmulhuw    xmm0, xmm2       // rgb * a
-    movdqa     xmm1, [eax]      // read 4 pixels
+    movdqu     xmm1, [eax]      // read 4 pixels
     punpckhbw  xmm1, xmm1       // next 2 pixels
     pshufhw    xmm2, xmm1, 0FFh // 8 alpha words
     pshuflw    xmm2, xmm2, 0FFh
     pmulhuw    xmm1, xmm2       // rgb * a
-    movdqa     xmm2, [eax]      // alphas
+    movdqu     xmm2, [eax]      // alphas
     lea        eax, [eax + 16]
     psrlw      xmm0, 8
     pand       xmm2, xmm4
@@ -4874,9 +3786,9 @@ void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width) {
     packuswb   xmm0, xmm1
     pand       xmm0, xmm5       // keep original alphas
     por        xmm0, xmm2
-    sub        ecx, 4
-    movdqa     [edx], xmm0
+    movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 4
     jg         convertloop
 
     ret
@@ -4904,7 +3816,6 @@ void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
     movdqa     xmm4, kShuffleAlpha0
     movdqa     xmm5, kShuffleAlpha1
 
-    align      4
  convertloop:
     movdqu     xmm0, [eax]      // read 4 pixels
     pshufb     xmm0, xmm4       // isolate first 2 alphas
@@ -4923,9 +3834,9 @@ void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
     psrlw      xmm1, 8
     packuswb   xmm0, xmm1
     por        xmm0, xmm2       // copy original alpha
-    sub        ecx, 4
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 4
     jg         convertloop
 
     ret
@@ -4935,11 +3846,8 @@ void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
 
 #ifdef HAS_ARGBATTENUATEROW_AVX2
 // Shuffle table duplicating alpha.
-static const ulvec8 kShuffleAlpha_AVX2 = {
-  6u, 7u, 6u, 7u, 6u, 7u, 128u, 128u,
-  14u, 15u, 14u, 15u, 14u, 15u, 128u, 128u,
-  6u, 7u, 6u, 7u, 6u, 7u, 128u, 128u,
-  14u, 15u, 14u, 15u, 14u, 15u, 128u, 128u,
+static const uvec8 kShuffleAlpha_AVX2 = {
+  6u, 7u, 6u, 7u, 6u, 7u, 128u, 128u, 14u, 15u, 14u, 15u, 14u, 15u, 128u, 128u
 };
 __declspec(naked) __declspec(align(16))
 void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width) {
@@ -4948,11 +3856,10 @@ void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width) {
     mov        edx, [esp + 8]   // dst_argb
     mov        ecx, [esp + 12]  // width
     sub        edx, eax
-    vmovdqa    ymm4, kShuffleAlpha_AVX2
+    vbroadcastf128 ymm4,kShuffleAlpha_AVX2
     vpcmpeqb   ymm5, ymm5, ymm5 // generate mask 0xff000000
     vpslld     ymm5, ymm5, 24
 
-    align      4
  convertloop:
     vmovdqu    ymm6, [eax]       // read 8 pixels.
     vpunpcklbw ymm0, ymm6, ymm6  // low 4 pixels. mutated.
@@ -4966,9 +3873,9 @@ void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width) {
     vpsrlw     ymm1, ymm1, 8
     vpackuswb  ymm0, ymm0, ymm1  // unmutated.
     vpor       ymm0, ymm0, ymm6  // copy original alpha
-    sub        ecx, 8
     vmovdqu    [eax + edx], ymm0
     lea        eax, [eax + 32]
+    sub        ecx, 8
     jg         convertloop
 
     vzeroupper
@@ -4979,7 +3886,6 @@ void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width) {
 
 #ifdef HAS_ARGBUNATTENUATEROW_SSE2
 // Unattenuate 4 pixels at a time.
-// Aligned to 16 bytes.
 __declspec(naked) __declspec(align(16))
 void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb,
                              int width) {
@@ -4990,7 +3896,6 @@ void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb,
     mov        edx, [esp + 8 + 8]   // dst_argb
     mov        ecx, [esp + 8 + 12]  // width
 
-    align      4
  convertloop:
     movdqu     xmm0, [eax]      // read 4 pixels
     movzx      esi, byte ptr [eax + 3]  // first alpha
@@ -5016,9 +3921,9 @@ void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb,
     lea        eax, [eax + 16]
 
     packuswb   xmm0, xmm1
-    sub        ecx, 4
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 4
     jg         convertloop
     pop        edi
     pop        esi
@@ -5029,9 +3934,8 @@ void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb,
 
 #ifdef HAS_ARGBUNATTENUATEROW_AVX2
 // Shuffle table duplicating alpha.
-static const ulvec8 kUnattenShuffleAlpha_AVX2 = {
-  0u, 1u, 0u, 1u, 0u, 1u, 6u, 7u, 8u, 9u, 8u, 9u, 8u, 9u, 14u, 15,
-  0u, 1u, 0u, 1u, 0u, 1u, 6u, 7u, 8u, 9u, 8u, 9u, 8u, 9u, 14u, 15,
+static const uvec8 kUnattenShuffleAlpha_AVX2 = {
+  0u, 1u, 0u, 1u, 0u, 1u, 6u, 7u, 8u, 9u, 8u, 9u, 8u, 9u, 14u, 15u
 };
 // TODO(fbarchard): Enable USE_GATHER for future hardware if faster.
 // USE_GATHER is not on by default, due to being a slow instruction.
@@ -5044,9 +3948,8 @@ void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb,
     mov        edx, [esp + 8]   // dst_argb
     mov        ecx, [esp + 12]  // width
     sub        edx, eax
-    vmovdqa    ymm4, kUnattenShuffleAlpha_AVX2
+    vbroadcastf128 ymm4, kUnattenShuffleAlpha_AVX2
 
-    align      4
  convertloop:
     vmovdqu    ymm6, [eax]       // read 8 pixels.
     vpcmpeqb   ymm5, ymm5, ymm5  // generate mask 0xffffffff for gather.
@@ -5061,9 +3964,9 @@ void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb,
     vpmulhuw   ymm0, ymm0, ymm2  // rgb * ia
     vpmulhuw   ymm1, ymm1, ymm3  // rgb * ia
     vpackuswb  ymm0, ymm0, ymm1  // unmutated.
-    sub        ecx, 8
     vmovdqu    [eax + edx], ymm0
     lea        eax, [eax + 32]
+    sub        ecx, 8
     jg         convertloop
 
     vzeroupper
@@ -5080,12 +3983,11 @@ void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb,
     mov        edx, [esp + 8]   // dst_argb
     mov        ecx, [esp + 12]  // width
     sub        edx, eax
-    vmovdqa    ymm5, kUnattenShuffleAlpha_AVX2
+    vbroadcastf128 ymm5, kUnattenShuffleAlpha_AVX2
 
     push       esi
     push       edi
 
-    align      4
  convertloop:
     // replace VPGATHER
     movzx      esi, byte ptr [eax + 3]                 // alpha0
@@ -5123,9 +4025,9 @@ void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb,
     vpmulhuw   ymm0, ymm0, ymm2  // rgb * ia
     vpmulhuw   ymm1, ymm1, ymm3  // rgb * ia
     vpackuswb  ymm0, ymm0, ymm1  // unmutated.
-    sub        ecx, 8
     vmovdqu    [eax + edx], ymm0
     lea        eax, [eax + 32]
+    sub        ecx, 8
     jg         convertloop
 
     pop        edi
@@ -5148,18 +4050,17 @@ void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
     movdqa     xmm4, kARGBToYJ
     movdqa     xmm5, kAddYJ64
 
-    align      4
  convertloop:
-    movdqa     xmm0, [eax]  // G
-    movdqa     xmm1, [eax + 16]
+    movdqu     xmm0, [eax]  // G
+    movdqu     xmm1, [eax + 16]
     pmaddubsw  xmm0, xmm4
     pmaddubsw  xmm1, xmm4
     phaddw     xmm0, xmm1
     paddw      xmm0, xmm5  // Add .5 for rounding.
     psrlw      xmm0, 7
     packuswb   xmm0, xmm0   // 8 G bytes
-    movdqa     xmm2, [eax]  // A
-    movdqa     xmm3, [eax + 16]
+    movdqu     xmm2, [eax]  // A
+    movdqu     xmm3, [eax + 16]
     lea        eax, [eax + 32]
     psrld      xmm2, 24
     psrld      xmm3, 24
@@ -5171,10 +4072,10 @@ void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
     movdqa     xmm1, xmm0
     punpcklwd  xmm0, xmm3   // GGGA first 4
     punpckhwd  xmm1, xmm3   // GGGA next 4
-    sub        ecx, 8
-    movdqa     [edx], xmm0
-    movdqa     [edx + 16], xmm1
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
     lea        edx, [edx + 32]
+    sub        ecx, 8
     jg         convertloop
     ret
   }
@@ -5208,32 +4109,31 @@ void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width) {
     movdqa     xmm3, kARGBToSepiaG
     movdqa     xmm4, kARGBToSepiaR
 
-    align      4
  convertloop:
-    movdqa     xmm0, [eax]  // B
-    movdqa     xmm6, [eax + 16]
+    movdqu     xmm0, [eax]  // B
+    movdqu     xmm6, [eax + 16]
     pmaddubsw  xmm0, xmm2
     pmaddubsw  xmm6, xmm2
     phaddw     xmm0, xmm6
     psrlw      xmm0, 7
     packuswb   xmm0, xmm0   // 8 B values
-    movdqa     xmm5, [eax]  // G
-    movdqa     xmm1, [eax + 16]
+    movdqu     xmm5, [eax]  // G
+    movdqu     xmm1, [eax + 16]
     pmaddubsw  xmm5, xmm3
     pmaddubsw  xmm1, xmm3
     phaddw     xmm5, xmm1
     psrlw      xmm5, 7
     packuswb   xmm5, xmm5   // 8 G values
     punpcklbw  xmm0, xmm5   // 8 BG values
-    movdqa     xmm5, [eax]  // R
-    movdqa     xmm1, [eax + 16]
+    movdqu     xmm5, [eax]  // R
+    movdqu     xmm1, [eax + 16]
     pmaddubsw  xmm5, xmm4
     pmaddubsw  xmm1, xmm4
     phaddw     xmm5, xmm1
     psrlw      xmm5, 7
     packuswb   xmm5, xmm5   // 8 R values
-    movdqa     xmm6, [eax]  // A
-    movdqa     xmm1, [eax + 16]
+    movdqu     xmm6, [eax]  // A
+    movdqu     xmm1, [eax + 16]
     psrld      xmm6, 24
     psrld      xmm1, 24
     packuswb   xmm6, xmm1
@@ -5242,10 +4142,10 @@ void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width) {
     movdqa     xmm1, xmm0   // Weave BG, RA together
     punpcklwd  xmm0, xmm5   // BGRA first 4
     punpckhwd  xmm1, xmm5   // BGRA next 4
-    sub        ecx, 8
-    movdqa     [eax], xmm0
-    movdqa     [eax + 16], xmm1
+    movdqu     [eax], xmm0
+    movdqu     [eax + 16], xmm1
     lea        eax, [eax + 32]
+    sub        ecx, 8
     jg         convertloop
     ret
   }
@@ -5271,14 +4171,13 @@ void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
     pshufd     xmm5, xmm5, 0xff
     mov        ecx, [esp + 16]  /* width */
 
-    align      4
  convertloop:
-    movdqa     xmm0, [eax]  // B
-    movdqa     xmm7, [eax + 16]
+    movdqu     xmm0, [eax]  // B
+    movdqu     xmm7, [eax + 16]
     pmaddubsw  xmm0, xmm2
     pmaddubsw  xmm7, xmm2
-    movdqa     xmm6, [eax]  // G
-    movdqa     xmm1, [eax + 16]
+    movdqu     xmm6, [eax]  // G
+    movdqu     xmm1, [eax + 16]
     pmaddubsw  xmm6, xmm3
     pmaddubsw  xmm1, xmm3
     phaddsw    xmm0, xmm7   // B
@@ -5288,13 +4187,13 @@ void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
     packuswb   xmm0, xmm0   // 8 B values
     packuswb   xmm6, xmm6   // 8 G values
     punpcklbw  xmm0, xmm6   // 8 BG values
-    movdqa     xmm1, [eax]  // R
-    movdqa     xmm7, [eax + 16]
+    movdqu     xmm1, [eax]  // R
+    movdqu     xmm7, [eax + 16]
     pmaddubsw  xmm1, xmm4
     pmaddubsw  xmm7, xmm4
     phaddsw    xmm1, xmm7   // R
-    movdqa     xmm6, [eax]  // A
-    movdqa     xmm7, [eax + 16]
+    movdqu     xmm6, [eax]  // A
+    movdqu     xmm7, [eax + 16]
     pmaddubsw  xmm6, xmm5
     pmaddubsw  xmm7, xmm5
     phaddsw    xmm6, xmm7   // A
@@ -5306,11 +4205,11 @@ void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
     movdqa     xmm6, xmm0   // Weave BG, RA together
     punpcklwd  xmm0, xmm1   // BGRA first 4
     punpckhwd  xmm6, xmm1   // BGRA next 4
-    sub        ecx, 8
-    movdqa     [edx], xmm0
-    movdqa     [edx + 16], xmm6
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm6
     lea        eax, [eax + 32]
     lea        edx, [edx + 32]
+    sub        ecx, 8
     jg         convertloop
     ret
   }
@@ -5319,7 +4218,6 @@ void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
 
 #ifdef HAS_ARGBQUANTIZEROW_SSE2
 // Quantize 4 ARGB pixels (16 bytes).
-// Aligned to 16 bytes.
 __declspec(naked) __declspec(align(16))
 void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
                           int interval_offset, int width) {
@@ -5339,25 +4237,24 @@ void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
     pcmpeqb    xmm6, xmm6  // generate mask 0xff000000
     pslld      xmm6, 24
 
-    align      4
  convertloop:
-    movdqa     xmm0, [eax]  // read 4 pixels
+    movdqu     xmm0, [eax]  // read 4 pixels
     punpcklbw  xmm0, xmm5   // first 2 pixels
     pmulhuw    xmm0, xmm2   // pixel * scale >> 16
-    movdqa     xmm1, [eax]  // read 4 pixels
+    movdqu     xmm1, [eax]  // read 4 pixels
     punpckhbw  xmm1, xmm5   // next 2 pixels
     pmulhuw    xmm1, xmm2
     pmullw     xmm0, xmm3   // * interval_size
-    movdqa     xmm7, [eax]  // read 4 pixels
+    movdqu     xmm7, [eax]  // read 4 pixels
     pmullw     xmm1, xmm3
     pand       xmm7, xmm6   // mask alpha
     paddw      xmm0, xmm4   // + interval_size / 2
     paddw      xmm1, xmm4
     packuswb   xmm0, xmm1
     por        xmm0, xmm7
-    sub        ecx, 4
-    movdqa     [eax], xmm0
+    movdqu     [eax], xmm0
     lea        eax, [eax + 16]
+    sub        ecx, 4
     jg         convertloop
     ret
   }
@@ -5366,7 +4263,6 @@ void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
 
 #ifdef HAS_ARGBSHADEROW_SSE2
 // Shade 4 pixels at a time by specified value.
-// Aligned to 16 bytes.
 __declspec(naked) __declspec(align(16))
 void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
                        uint32 value) {
@@ -5378,9 +4274,8 @@ void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
     punpcklbw  xmm2, xmm2
     punpcklqdq xmm2, xmm2
 
-    align      4
  convertloop:
-    movdqa     xmm0, [eax]      // read 4 pixels
+    movdqu     xmm0, [eax]      // read 4 pixels
     lea        eax, [eax + 16]
     movdqa     xmm1, xmm0
     punpcklbw  xmm0, xmm0       // first 2
@@ -5390,9 +4285,9 @@ void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
     psrlw      xmm0, 8
     psrlw      xmm1, 8
     packuswb   xmm0, xmm1
-    sub        ecx, 4
-    movdqa     [edx], xmm0
+    movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 4
     jg         convertloop
 
     ret
@@ -5413,7 +4308,6 @@ void ARGBMultiplyRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
     mov        ecx, [esp + 4 + 16]  // width
     pxor       xmm5, xmm5  // constant 0
 
-    align      4
  convertloop:
     movdqu     xmm0, [eax]        // read 4 pixels from src_argb0
     movdqu     xmm2, [esi]        // read 4 pixels from src_argb1
@@ -5428,9 +4322,9 @@ void ARGBMultiplyRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
     lea        eax, [eax + 16]
     lea        esi, [esi + 16]
     packuswb   xmm0, xmm1
-    sub        ecx, 4
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 4
     jg         convertloop
 
     pop        esi
@@ -5455,16 +4349,15 @@ void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
     sub        ecx, 4
     jl         convertloop49
 
-    align      4
  convertloop4:
     movdqu     xmm0, [eax]        // read 4 pixels from src_argb0
     lea        eax, [eax + 16]
     movdqu     xmm1, [esi]        // read 4 pixels from src_argb1
     lea        esi, [esi + 16]
     paddusb    xmm0, xmm1         // src_argb0 + src_argb1
-    sub        ecx, 4
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 4
     jge        convertloop4
 
  convertloop49:
@@ -5477,9 +4370,9 @@ void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
     movd       xmm1, [esi]        // read 1 pixels from src_argb1
     lea        esi, [esi + 4]
     paddusb    xmm0, xmm1         // src_argb0 + src_argb1
-    sub        ecx, 1
     movd       [edx], xmm0
     lea        edx, [edx + 4]
+    sub        ecx, 1
     jge        convertloop1
 
  convertloop19:
@@ -5501,16 +4394,15 @@ void ARGBSubtractRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
     mov        edx, [esp + 4 + 12]  // dst_argb
     mov        ecx, [esp + 4 + 16]  // width
 
-    align      4
  convertloop:
     movdqu     xmm0, [eax]        // read 4 pixels from src_argb0
     lea        eax, [eax + 16]
     movdqu     xmm1, [esi]        // read 4 pixels from src_argb1
     lea        esi, [esi + 16]
     psubusb    xmm0, xmm1         // src_argb0 - src_argb1
-    sub        ecx, 4
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 4
     jg         convertloop
 
     pop        esi
@@ -5532,7 +4424,6 @@ void ARGBMultiplyRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
     mov        ecx, [esp + 4 + 16]  // width
     vpxor      ymm5, ymm5, ymm5     // constant 0
 
-    align      4
  convertloop:
     vmovdqu    ymm1, [eax]        // read 8 pixels from src_argb0
     lea        eax, [eax + 32]
@@ -5569,7 +4460,6 @@ void ARGBAddRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
     mov        edx, [esp + 4 + 12]  // dst_argb
     mov        ecx, [esp + 4 + 16]  // width
 
-    align      4
  convertloop:
     vmovdqu    ymm0, [eax]              // read 8 pixels from src_argb0
     lea        eax, [eax + 32]
@@ -5599,7 +4489,6 @@ void ARGBSubtractRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
     mov        edx, [esp + 4 + 12]  // dst_argb
     mov        ecx, [esp + 4 + 16]  // width
 
-    align      4
  convertloop:
     vmovdqu    ymm0, [eax]              // read 8 pixels from src_argb0
     lea        eax, [eax + 32]
@@ -5638,7 +4527,6 @@ void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
     sub        edx, eax
     pxor       xmm5, xmm5  // constant 0
 
-    align      4
  convertloop:
     movq       xmm0, qword ptr [eax]            // read 8 pixels from src_y0[0]
     movq       xmm1, qword ptr [eax + 2]        // read 8 pixels from src_y0[2]
@@ -5662,9 +4550,9 @@ void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
     psubw      xmm1, xmm0
     pmaxsw     xmm0, xmm1
     packuswb   xmm0, xmm0
-    sub        ecx, 8
     movq       qword ptr [eax + edx], xmm0
     lea        eax, [eax + 8]
+    sub        ecx, 8
     jg         convertloop
 
     pop        edi
@@ -5692,7 +4580,6 @@ void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
     sub        edx, eax
     pxor       xmm5, xmm5  // constant 0
 
-    align      4
  convertloop:
     movq       xmm0, qword ptr [eax]            // read 8 pixels from src_y0[0]
     movq       xmm1, qword ptr [eax + esi]      // read 8 pixels from src_y1[0]
@@ -5716,9 +4603,9 @@ void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
     psubw      xmm1, xmm0
     pmaxsw     xmm0, xmm1
     packuswb   xmm0, xmm0
-    sub        ecx, 8
     movq       qword ptr [eax + edx], xmm0
     lea        eax, [eax + 8]
+    sub        ecx, 8
     jg         convertloop
 
     pop        esi
@@ -5746,10 +4633,9 @@ void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
     pcmpeqb    xmm5, xmm5           // alpha 255
     pslld      xmm5, 24             // 0xff000000
 
-    align      4
  convertloop:
-    movdqa     xmm0, [eax]            // read 16 pixels src_sobelx
-    movdqa     xmm1, [eax + esi]      // read 16 pixels src_sobely
+    movdqu     xmm0, [eax]            // read 16 pixels src_sobelx
+    movdqu     xmm1, [eax + esi]      // read 16 pixels src_sobely
     lea        eax, [eax + 16]
     paddusb    xmm0, xmm1             // sobel = sobelx + sobely
     movdqa     xmm2, xmm0             // GG
@@ -5765,12 +4651,12 @@ void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
     punpckhwd  xmm0, xmm0             // Last 4
     por        xmm3, xmm5             // GGGA
     por        xmm0, xmm5
-    sub        ecx, 16
-    movdqa     [edx], xmm1
-    movdqa     [edx + 16], xmm2
-    movdqa     [edx + 32], xmm3
-    movdqa     [edx + 48], xmm0
+    movdqu     [edx], xmm1
+    movdqu     [edx + 16], xmm2
+    movdqu     [edx + 32], xmm3
+    movdqu     [edx + 48], xmm0
     lea        edx, [edx + 64]
+    sub        ecx, 16
     jg         convertloop
 
     pop        esi
@@ -5792,15 +4678,14 @@ void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
     mov        ecx, [esp + 4 + 16]  // width
     sub        esi, eax
 
-    align      4
  convertloop:
-    movdqa     xmm0, [eax]            // read 16 pixels src_sobelx
-    movdqa     xmm1, [eax + esi]      // read 16 pixels src_sobely
+    movdqu     xmm0, [eax]            // read 16 pixels src_sobelx
+    movdqu     xmm1, [eax + esi]      // read 16 pixels src_sobely
     lea        eax, [eax + 16]
     paddusb    xmm0, xmm1             // sobel = sobelx + sobely
-    sub        ecx, 16
-    movdqa     [edx], xmm0
+    movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 16
     jg         convertloop
 
     pop        esi
@@ -5827,10 +4712,9 @@ void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
     sub        esi, eax
     pcmpeqb    xmm5, xmm5           // alpha 255
 
-    align      4
  convertloop:
-    movdqa     xmm0, [eax]            // read 16 pixels src_sobelx
-    movdqa     xmm1, [eax + esi]      // read 16 pixels src_sobely
+    movdqu     xmm0, [eax]            // read 16 pixels src_sobelx
+    movdqu     xmm1, [eax + esi]      // read 16 pixels src_sobely
     lea        eax, [eax + 16]
     movdqa     xmm2, xmm0
     paddusb    xmm2, xmm1             // sobel = sobelx + sobely
@@ -5846,12 +4730,12 @@ void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
     movdqa     xmm7, xmm1             // YSXA
     punpcklwd  xmm7, xmm0             // Next 4
     punpckhwd  xmm1, xmm0             // Last 4
-    sub        ecx, 16
-    movdqa     [edx], xmm6
-    movdqa     [edx + 16], xmm4
-    movdqa     [edx + 32], xmm7
-    movdqa     [edx + 48], xmm1
+    movdqu     [edx], xmm6
+    movdqu     [edx + 16], xmm4
+    movdqu     [edx + 32], xmm7
+    movdqu     [edx + 48], xmm1
     lea        edx, [edx + 64]
+    sub        ecx, 16
     jg         convertloop
 
     pop        esi
@@ -5872,8 +4756,7 @@ void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
 // area is the number of pixels in the area being averaged.
 // dst points to pixel to store result to.
 // count is number of averaged pixels to produce.
-// Does 4 pixels at a time, requires CumulativeSum pointers to be 16 byte
-// aligned.
+// Does 4 pixels at a time.
 void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
                                     int width, int area, uint8* dst,
                                     int count) {
@@ -5903,13 +4786,12 @@ void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
     packssdw   xmm5, xmm5           // 16 bit shorts
 
     // 4 pixel loop small blocks.
-    align      4
   s4:
     // top left
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    movdqa     xmm2, [eax + 32]
-    movdqa     xmm3, [eax + 48]
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
 
     // - top right
     psubd      xmm0, [eax + edx * 4]
@@ -5946,13 +4828,12 @@ void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
     jmp        l4b
 
     // 4 pixel loop
-    align      4
   l4:
     // top left
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    movdqa     xmm2, [eax + 32]
-    movdqa     xmm3, [eax + 48]
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
 
     // - top right
     psubd      xmm0, [eax + edx * 4]
@@ -5999,9 +4880,8 @@ void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
     jl         l1b
 
     // 1 pixel loop
-    align      4
   l1:
-    movdqa     xmm0, [eax]
+    movdqu     xmm0, [eax]
     psubd      xmm0, [eax + edx * 4]
     lea        eax, [eax + 16]
     psubd      xmm0, [esi]
@@ -6040,7 +4920,6 @@ void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
     jne        l4b
 
     // 4 pixel loop
-    align      4
   l4:
     movdqu     xmm2, [eax]  // 4 argb pixels 16 bytes.
     lea        eax, [eax + 16]
@@ -6057,26 +4936,26 @@ void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
     punpckhwd  xmm5, xmm1
 
     paddd      xmm0, xmm2
-    movdqa     xmm2, [esi]  // previous row above.
+    movdqu     xmm2, [esi]  // previous row above.
     paddd      xmm2, xmm0
 
     paddd      xmm0, xmm3
-    movdqa     xmm3, [esi + 16]
+    movdqu     xmm3, [esi + 16]
     paddd      xmm3, xmm0
 
     paddd      xmm0, xmm4
-    movdqa     xmm4, [esi + 32]
+    movdqu     xmm4, [esi + 32]
     paddd      xmm4, xmm0
 
     paddd      xmm0, xmm5
-    movdqa     xmm5, [esi + 48]
+    movdqu     xmm5, [esi + 48]
     lea        esi, [esi + 64]
     paddd      xmm5, xmm0
 
-    movdqa     [edx], xmm2
-    movdqa     [edx + 16], xmm3
-    movdqa     [edx + 32], xmm4
-    movdqa     [edx + 48], xmm5
+    movdqu     [edx], xmm2
+    movdqu     [edx + 16], xmm3
+    movdqu     [edx + 32], xmm4
+    movdqu     [edx + 48], xmm5
 
     lea        edx, [edx + 64]
     sub        ecx, 4
@@ -6087,7 +4966,6 @@ void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
     jl         l1b
 
     // 1 pixel loop
-    align      4
   l1:
     movd       xmm2, dword ptr [eax]  // 1 argb pixel 4 bytes.
     lea        eax, [eax + 4]
@@ -6142,7 +5020,6 @@ void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
     addps      xmm4, xmm4    // dudv *= 4
 
     // 4 pixel loop
-    align      4
   l4:
     cvttps2dq  xmm0, xmm2    // x, y float to int first 2
     cvttps2dq  xmm1, xmm3    // x, y float to int next 2
@@ -6164,9 +5041,9 @@ void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
     movd       xmm0, [eax + edi]  // read pixel 3
     punpckldq  xmm6, xmm0     // combine pixel 2 and 3
     addps      xmm3, xmm4    // x, y += dx, dy next 2
-    sub        ecx, 4
     movq       qword ptr 8[edx], xmm6
     lea        edx, [edx + 16]
+    sub        ecx, 4
     jge        l4
 
   l4b:
@@ -6174,7 +5051,6 @@ void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
     jl         l1b
 
     // 1 pixel loop
-    align      4
   l1:
     cvttps2dq  xmm0, xmm2    // x, y float to int
     packssdw   xmm0, xmm0    // x, y as shorts
@@ -6182,9 +5058,9 @@ void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
     addps      xmm2, xmm7    // x, y += dx, dy
     movd       esi, xmm0
     movd       xmm0, [eax + esi]  // copy a pixel
-    sub        ecx, 1
     movd       [edx], xmm0
     lea        edx, [edx + 4]
+    sub        ecx, 1
     jge        l1
   l1b:
     pop        edi
@@ -6195,11 +5071,11 @@ void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
 #endif  // HAS_ARGBAFFINEROW_SSE2
 
 #ifdef HAS_INTERPOLATEROW_AVX2
-// Bilinear filter 16x2 -> 16x1
+// Bilinear filter 32x2 -> 32x1
 __declspec(naked) __declspec(align(16))
 void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
-                          ptrdiff_t src_stride, int dst_width,
-                          int source_y_fraction) {
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) {
   __asm {
     push       esi
     push       edi
@@ -6229,7 +5105,6 @@ void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
     vpxor      ymm0, ymm0, ymm0
     vpermd     ymm5, ymm0, ymm5
 
-    align      4
   xloop:
     vmovdqu    ymm0, [esi]
     vmovdqu    ymm2, [esi + edx]
@@ -6240,51 +5115,49 @@ void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
     vpsrlw     ymm0, ymm0, 7
     vpsrlw     ymm1, ymm1, 7
     vpackuswb  ymm0, ymm0, ymm1  // unmutates
-    sub        ecx, 32
-    vmovdqu    [esi + edi], ymm0
-    lea        esi, [esi + 32]
-    jg         xloop
-    jmp        xloop99
-
-    // Blend 25 / 75.
-    align      4
-  xloop25:
-    vmovdqu    ymm0, [esi]
-    vpavgb     ymm0, ymm0, [esi + edx]
-    vpavgb     ymm0, ymm0, [esi + edx]
-    sub        ecx, 32
-    vmovdqu    [esi + edi], ymm0
-    lea        esi, [esi + 32]
-    jg         xloop25
-    jmp        xloop99
-
-    // Blend 50 / 50.
-    align      4
-  xloop50:
-    vmovdqu    ymm0, [esi]
-    vpavgb     ymm0, ymm0, [esi + edx]
-    sub        ecx, 32
     vmovdqu    [esi + edi], ymm0
     lea        esi, [esi + 32]
-    jg         xloop50
-    jmp        xloop99
-
-    // Blend 75 / 25.
-    align      4
-  xloop75:
-    vmovdqu    ymm0, [esi + edx]
-    vpavgb     ymm0, ymm0, [esi]
-    vpavgb     ymm0, ymm0, [esi]
     sub        ecx, 32
-    vmovdqu     [esi + edi], ymm0
-    lea        esi, [esi + 32]
-    jg         xloop75
+    jg         xloop
     jmp        xloop99
 
-    // Blend 100 / 0 - Copy row unchanged.
-    align      4
-  xloop100:
-    rep movsb
+   // Blend 25 / 75.
+ xloop25:
+   vmovdqu    ymm0, [esi]
+   vmovdqu    ymm1, [esi + edx]
+   vpavgb     ymm0, ymm0, ymm1
+   vpavgb     ymm0, ymm0, ymm1
+   vmovdqu    [esi + edi], ymm0
+   lea        esi, [esi + 32]
+   sub        ecx, 32
+   jg         xloop25
+   jmp        xloop99
+
+   // Blend 50 / 50.
+ xloop50:
+   vmovdqu    ymm0, [esi]
+   vpavgb     ymm0, ymm0, [esi + edx]
+   vmovdqu    [esi + edi], ymm0
+   lea        esi, [esi + 32]
+   sub        ecx, 32
+   jg         xloop50
+   jmp        xloop99
+
+   // Blend 75 / 25.
+ xloop75:
+   vmovdqu    ymm1, [esi]
+   vmovdqu    ymm0, [esi + edx]
+   vpavgb     ymm0, ymm0, ymm1
+   vpavgb     ymm0, ymm0, ymm1
+   vmovdqu    [esi + edi], ymm0
+   lea        esi, [esi + 32]
+   sub        ecx, 32
+   jg         xloop75
+   jmp        xloop99
+
+   // Blend 100 / 0 - Copy row unchanged.
+ xloop100:
+   rep movsb
 
   xloop99:
     pop        edi
@@ -6295,7 +5168,6 @@ void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
 }
 #endif  // HAS_INTERPOLATEROW_AVX2
 
-#ifdef HAS_INTERPOLATEROW_SSSE3
 // Bilinear filter 16x2 -> 16x1
 __declspec(naked) __declspec(align(16))
 void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
@@ -6329,226 +5201,6 @@ void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
     punpcklwd  xmm5, xmm5
     pshufd     xmm5, xmm5, 0
 
-    align      4
-  xloop:
-    movdqa     xmm0, [esi]
-    movdqa     xmm2, [esi + edx]
-    movdqa     xmm1, xmm0
-    punpcklbw  xmm0, xmm2
-    punpckhbw  xmm1, xmm2
-    pmaddubsw  xmm0, xmm5
-    pmaddubsw  xmm1, xmm5
-    psrlw      xmm0, 7
-    psrlw      xmm1, 7
-    packuswb   xmm0, xmm1
-    sub        ecx, 16
-    movdqa     [esi + edi], xmm0
-    lea        esi, [esi + 16]
-    jg         xloop
-    jmp        xloop99
-
-    // Blend 25 / 75.
-    align      4
-  xloop25:
-    movdqa     xmm0, [esi]
-    movdqa     xmm1, [esi + edx]
-    pavgb      xmm0, xmm1
-    pavgb      xmm0, xmm1
-    sub        ecx, 16
-    movdqa     [esi + edi], xmm0
-    lea        esi, [esi + 16]
-    jg         xloop25
-    jmp        xloop99
-
-    // Blend 50 / 50.
-    align      4
-  xloop50:
-    movdqa     xmm0, [esi]
-    movdqa     xmm1, [esi + edx]
-    pavgb      xmm0, xmm1
-    sub        ecx, 16
-    movdqa     [esi + edi], xmm0
-    lea        esi, [esi + 16]
-    jg         xloop50
-    jmp        xloop99
-
-    // Blend 75 / 25.
-    align      4
-  xloop75:
-    movdqa     xmm1, [esi]
-    movdqa     xmm0, [esi + edx]
-    pavgb      xmm0, xmm1
-    pavgb      xmm0, xmm1
-    sub        ecx, 16
-    movdqa     [esi + edi], xmm0
-    lea        esi, [esi + 16]
-    jg         xloop75
-    jmp        xloop99
-
-    // Blend 100 / 0 - Copy row unchanged.
-    align      4
-  xloop100:
-    movdqa     xmm0, [esi]
-    sub        ecx, 16
-    movdqa     [esi + edi], xmm0
-    lea        esi, [esi + 16]
-    jg         xloop100
-
-  xloop99:
-    pop        edi
-    pop        esi
-    ret
-  }
-}
-#endif  // HAS_INTERPOLATEROW_SSSE3
-
-#ifdef HAS_INTERPOLATEROW_SSE2
-// Bilinear filter 16x2 -> 16x1
-__declspec(naked) __declspec(align(16))
-void InterpolateRow_SSE2(uint8* dst_ptr, const uint8* src_ptr,
-                         ptrdiff_t src_stride, int dst_width,
-                         int source_y_fraction) {
-  __asm {
-    push       esi
-    push       edi
-    mov        edi, [esp + 8 + 4]   // dst_ptr
-    mov        esi, [esp + 8 + 8]   // src_ptr
-    mov        edx, [esp + 8 + 12]  // src_stride
-    mov        ecx, [esp + 8 + 16]  // dst_width
-    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
-    sub        edi, esi
-    // Dispatch to specialized filters if applicable.
-    cmp        eax, 0
-    je         xloop100  // 0 / 256.  Blend 100 / 0.
-    cmp        eax, 64
-    je         xloop75   // 64 / 256 is 0.25.  Blend 75 / 25.
-    cmp        eax, 128
-    je         xloop50   // 128 / 256 is 0.50.  Blend 50 / 50.
-    cmp        eax, 192
-    je         xloop25   // 192 / 256 is 0.75.  Blend 25 / 75.
-
-    movd       xmm5, eax            // xmm5 = y fraction
-    punpcklbw  xmm5, xmm5
-    psrlw      xmm5, 1
-    punpcklwd  xmm5, xmm5
-    punpckldq  xmm5, xmm5
-    punpcklqdq xmm5, xmm5
-    pxor       xmm4, xmm4
-
-    align      4
-  xloop:
-    movdqa     xmm0, [esi]  // row0
-    movdqa     xmm2, [esi + edx]  // row1
-    movdqa     xmm1, xmm0
-    movdqa     xmm3, xmm2
-    punpcklbw  xmm2, xmm4
-    punpckhbw  xmm3, xmm4
-    punpcklbw  xmm0, xmm4
-    punpckhbw  xmm1, xmm4
-    psubw      xmm2, xmm0  // row1 - row0
-    psubw      xmm3, xmm1
-    paddw      xmm2, xmm2  // 9 bits * 15 bits = 8.16
-    paddw      xmm3, xmm3
-    pmulhw     xmm2, xmm5  // scale diff
-    pmulhw     xmm3, xmm5
-    paddw      xmm0, xmm2  // sum rows
-    paddw      xmm1, xmm3
-    packuswb   xmm0, xmm1
-    sub        ecx, 16
-    movdqa     [esi + edi], xmm0
-    lea        esi, [esi + 16]
-    jg         xloop
-    jmp        xloop99
-
-    // Blend 25 / 75.
-    align      4
-  xloop25:
-    movdqa     xmm0, [esi]
-    movdqa     xmm1, [esi + edx]
-    pavgb      xmm0, xmm1
-    pavgb      xmm0, xmm1
-    sub        ecx, 16
-    movdqa     [esi + edi], xmm0
-    lea        esi, [esi + 16]
-    jg         xloop25
-    jmp        xloop99
-
-    // Blend 50 / 50.
-    align      4
-  xloop50:
-    movdqa     xmm0, [esi]
-    movdqa     xmm1, [esi + edx]
-    pavgb      xmm0, xmm1
-    sub        ecx, 16
-    movdqa     [esi + edi], xmm0
-    lea        esi, [esi + 16]
-    jg         xloop50
-    jmp        xloop99
-
-    // Blend 75 / 25.
-    align      4
-  xloop75:
-    movdqa     xmm1, [esi]
-    movdqa     xmm0, [esi + edx]
-    pavgb      xmm0, xmm1
-    pavgb      xmm0, xmm1
-    sub        ecx, 16
-    movdqa     [esi + edi], xmm0
-    lea        esi, [esi + 16]
-    jg         xloop75
-    jmp        xloop99
-
-    // Blend 100 / 0 - Copy row unchanged.
-    align      4
-  xloop100:
-    movdqa     xmm0, [esi]
-    sub        ecx, 16
-    movdqa     [esi + edi], xmm0
-    lea        esi, [esi + 16]
-    jg         xloop100
-
-  xloop99:
-    pop        edi
-    pop        esi
-    ret
-  }
-}
-#endif  // HAS_INTERPOLATEROW_SSE2
-
-// Bilinear filter 16x2 -> 16x1
-__declspec(naked) __declspec(align(16))
-void InterpolateRow_Unaligned_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
-                                    ptrdiff_t src_stride, int dst_width,
-                                    int source_y_fraction) {
-  __asm {
-    push       esi
-    push       edi
-    mov        edi, [esp + 8 + 4]   // dst_ptr
-    mov        esi, [esp + 8 + 8]   // src_ptr
-    mov        edx, [esp + 8 + 12]  // src_stride
-    mov        ecx, [esp + 8 + 16]  // dst_width
-    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
-    sub        edi, esi
-    shr        eax, 1
-    // Dispatch to specialized filters if applicable.
-    cmp        eax, 0
-    je         xloop100  // 0 / 128.  Blend 100 / 0.
-    cmp        eax, 32
-    je         xloop75   // 32 / 128 is 0.25.  Blend 75 / 25.
-    cmp        eax, 64
-    je         xloop50   // 64 / 128 is 0.50.  Blend 50 / 50.
-    cmp        eax, 96
-    je         xloop25   // 96 / 128 is 0.75.  Blend 25 / 75.
-
-    movd       xmm0, eax  // high fraction 0..127
-    neg        eax
-    add        eax, 128
-    movd       xmm5, eax  // low fraction 128..1
-    punpcklbw  xmm5, xmm0
-    punpcklwd  xmm5, xmm5
-    pshufd     xmm5, xmm5, 0
-
-    align      4
   xloop:
     movdqu     xmm0, [esi]
     movdqu     xmm2, [esi + edx]
@@ -6560,57 +5212,53 @@ void InterpolateRow_Unaligned_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
     psrlw      xmm0, 7
     psrlw      xmm1, 7
     packuswb   xmm0, xmm1
-    sub        ecx, 16
     movdqu     [esi + edi], xmm0
     lea        esi, [esi + 16]
+    sub        ecx, 16
     jg         xloop
     jmp        xloop99
 
     // Blend 25 / 75.
-    align      4
   xloop25:
     movdqu     xmm0, [esi]
     movdqu     xmm1, [esi + edx]
     pavgb      xmm0, xmm1
     pavgb      xmm0, xmm1
-    sub        ecx, 16
     movdqu     [esi + edi], xmm0
     lea        esi, [esi + 16]
+    sub        ecx, 16
     jg         xloop25
     jmp        xloop99
 
     // Blend 50 / 50.
-    align      4
   xloop50:
     movdqu     xmm0, [esi]
     movdqu     xmm1, [esi + edx]
     pavgb      xmm0, xmm1
-    sub        ecx, 16
     movdqu     [esi + edi], xmm0
     lea        esi, [esi + 16]
+    sub        ecx, 16
     jg         xloop50
     jmp        xloop99
 
     // Blend 75 / 25.
-    align      4
   xloop75:
     movdqu     xmm1, [esi]
     movdqu     xmm0, [esi + edx]
     pavgb      xmm0, xmm1
     pavgb      xmm0, xmm1
-    sub        ecx, 16
     movdqu     [esi + edi], xmm0
     lea        esi, [esi + 16]
+    sub        ecx, 16
     jg         xloop75
     jmp        xloop99
 
     // Blend 100 / 0 - Copy row unchanged.
-    align      4
   xloop100:
     movdqu     xmm0, [esi]
-    sub        ecx, 16
     movdqu     [esi + edi], xmm0
     lea        esi, [esi + 16]
+    sub        ecx, 16
     jg         xloop100
 
   xloop99:
@@ -6623,9 +5271,9 @@ void InterpolateRow_Unaligned_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
 #ifdef HAS_INTERPOLATEROW_SSE2
 // Bilinear filter 16x2 -> 16x1
 __declspec(naked) __declspec(align(16))
-void InterpolateRow_Unaligned_SSE2(uint8* dst_ptr, const uint8* src_ptr,
-                                   ptrdiff_t src_stride, int dst_width,
-                                   int source_y_fraction) {
+void InterpolateRow_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) {
   __asm {
     push       esi
     push       edi
@@ -6653,7 +5301,6 @@ void InterpolateRow_Unaligned_SSE2(uint8* dst_ptr, const uint8* src_ptr,
     punpcklqdq xmm5, xmm5
     pxor       xmm4, xmm4
 
-    align      4
   xloop:
     movdqu     xmm0, [esi]  // row0
     movdqu     xmm2, [esi + edx]  // row1
@@ -6672,57 +5319,53 @@ void InterpolateRow_Unaligned_SSE2(uint8* dst_ptr, const uint8* src_ptr,
     paddw      xmm0, xmm2  // sum rows
     paddw      xmm1, xmm3
     packuswb   xmm0, xmm1
-    sub        ecx, 16
     movdqu     [esi + edi], xmm0
     lea        esi, [esi + 16]
+    sub        ecx, 16
     jg         xloop
     jmp        xloop99
 
     // Blend 25 / 75.
-    align      4
   xloop25:
     movdqu     xmm0, [esi]
     movdqu     xmm1, [esi + edx]
     pavgb      xmm0, xmm1
     pavgb      xmm0, xmm1
-    sub        ecx, 16
     movdqu     [esi + edi], xmm0
     lea        esi, [esi + 16]
+    sub        ecx, 16
     jg         xloop25
     jmp        xloop99
 
     // Blend 50 / 50.
-    align      4
   xloop50:
     movdqu     xmm0, [esi]
     movdqu     xmm1, [esi + edx]
     pavgb      xmm0, xmm1
-    sub        ecx, 16
     movdqu     [esi + edi], xmm0
     lea        esi, [esi + 16]
+    sub        ecx, 16
     jg         xloop50
     jmp        xloop99
 
     // Blend 75 / 25.
-    align      4
   xloop75:
     movdqu     xmm1, [esi]
     movdqu     xmm0, [esi + edx]
     pavgb      xmm0, xmm1
     pavgb      xmm0, xmm1
-    sub        ecx, 16
     movdqu     [esi + edi], xmm0
     lea        esi, [esi + 16]
+    sub        ecx, 16
     jg         xloop75
     jmp        xloop99
 
     // Blend 100 / 0 - Copy row unchanged.
-    align      4
   xloop100:
     movdqu     xmm0, [esi]
-    sub        ecx, 16
     movdqu     [esi + edi], xmm0
     lea        esi, [esi + 16]
+    sub        ecx, 16
     jg         xloop100
 
   xloop99:
@@ -6733,84 +5376,6 @@ void InterpolateRow_Unaligned_SSE2(uint8* dst_ptr, const uint8* src_ptr,
 }
 #endif  // HAS_INTERPOLATEROW_SSE2
 
-__declspec(naked) __declspec(align(16))
-void HalfRow_SSE2(const uint8* src_uv, int src_uv_stride,
-                  uint8* dst_uv, int pix) {
-  __asm {
-    push       edi
-    mov        eax, [esp + 4 + 4]    // src_uv
-    mov        edx, [esp + 4 + 8]    // src_uv_stride
-    mov        edi, [esp + 4 + 12]   // dst_v
-    mov        ecx, [esp + 4 + 16]   // pix
-    sub        edi, eax
-
-    align      4
-  convertloop:
-    movdqa     xmm0, [eax]
-    pavgb      xmm0, [eax + edx]
-    sub        ecx, 16
-    movdqa     [eax + edi], xmm0
-    lea        eax,  [eax + 16]
-    jg         convertloop
-    pop        edi
-    ret
-  }
-}
-
-#ifdef HAS_HALFROW_AVX2
-__declspec(naked) __declspec(align(16))
-void HalfRow_AVX2(const uint8* src_uv, int src_uv_stride,
-                  uint8* dst_uv, int pix) {
-  __asm {
-    push       edi
-    mov        eax, [esp + 4 + 4]    // src_uv
-    mov        edx, [esp + 4 + 8]    // src_uv_stride
-    mov        edi, [esp + 4 + 12]   // dst_v
-    mov        ecx, [esp + 4 + 16]   // pix
-    sub        edi, eax
-
-    align      4
-  convertloop:
-    vmovdqu    ymm0, [eax]
-    vpavgb     ymm0, ymm0, [eax + edx]
-    sub        ecx, 32
-    vmovdqu    [eax + edi], ymm0
-    lea        eax,  [eax + 32]
-    jg         convertloop
-
-    pop        edi
-    vzeroupper
-    ret
-  }
-}
-#endif  // HAS_HALFROW_AVX2
-
-__declspec(naked) __declspec(align(16))
-void ARGBToBayerRow_SSSE3(const uint8* src_argb, uint8* dst_bayer,
-                          uint32 selector, int pix) {
-  __asm {
-    mov        eax, [esp + 4]    // src_argb
-    mov        edx, [esp + 8]    // dst_bayer
-    movd       xmm5, [esp + 12]  // selector
-    mov        ecx, [esp + 16]   // pix
-    pshufd     xmm5, xmm5, 0
-
-    align      4
-  wloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    lea        eax, [eax + 32]
-    pshufb     xmm0, xmm5
-    pshufb     xmm1, xmm5
-    punpckldq  xmm0, xmm1
-    sub        ecx, 8
-    movq       qword ptr [edx], xmm0
-    lea        edx, [edx + 8]
-    jg         wloop
-    ret
-  }
-}
-
 // Specialized ARGB to Bayer that just isolates G channel.
 __declspec(naked) __declspec(align(16))
 void ARGBToBayerGGRow_SSE2(const uint8* src_argb, uint8* dst_bayer,
@@ -6823,10 +5388,9 @@ void ARGBToBayerGGRow_SSE2(const uint8* src_argb, uint8* dst_bayer,
     pcmpeqb    xmm5, xmm5        // generate mask 0x000000ff
     psrld      xmm5, 24
 
-    align      4
   wloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
     lea        eax, [eax + 32]
     psrld      xmm0, 8  // Move green to bottom.
     psrld      xmm1, 8
@@ -6834,9 +5398,9 @@ void ARGBToBayerGGRow_SSE2(const uint8* src_argb, uint8* dst_bayer,
     pand       xmm1, xmm5
     packssdw   xmm0, xmm1
     packuswb   xmm0, xmm1
-    sub        ecx, 8
     movq       qword ptr [edx], xmm0
     lea        edx, [edx + 8]
+    sub        ecx, 8
     jg         wloop
     ret
   }
@@ -6850,46 +5414,19 @@ void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
     mov        eax, [esp + 4]    // src_argb
     mov        edx, [esp + 8]    // dst_argb
     mov        ecx, [esp + 12]   // shuffler
-    movdqa     xmm5, [ecx]
-    mov        ecx, [esp + 16]   // pix
-
-    align      4
-  wloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    lea        eax, [eax + 32]
-    pshufb     xmm0, xmm5
-    pshufb     xmm1, xmm5
-    sub        ecx, 8
-    movdqa     [edx], xmm0
-    movdqa     [edx + 16], xmm1
-    lea        edx, [edx + 32]
-    jg         wloop
-    ret
-  }
-}
-
-__declspec(naked) __declspec(align(16))
-void ARGBShuffleRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_argb,
-                                    const uint8* shuffler, int pix) {
-  __asm {
-    mov        eax, [esp + 4]    // src_argb
-    mov        edx, [esp + 8]    // dst_argb
-    mov        ecx, [esp + 12]   // shuffler
-    movdqa     xmm5, [ecx]
+    movdqu     xmm5, [ecx]
     mov        ecx, [esp + 16]   // pix
 
-    align      4
   wloop:
     movdqu     xmm0, [eax]
     movdqu     xmm1, [eax + 16]
     lea        eax, [eax + 32]
     pshufb     xmm0, xmm5
     pshufb     xmm1, xmm5
-    sub        ecx, 8
     movdqu     [edx], xmm0
     movdqu     [edx + 16], xmm1
     lea        edx, [edx + 32]
+    sub        ecx, 8
     jg         wloop
     ret
   }
@@ -6906,17 +5443,16 @@ void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb,
     vbroadcastf128 ymm5, [ecx]    // same shuffle in high as low.
     mov        ecx, [esp + 16]    // pix
 
-    align      4
   wloop:
     vmovdqu    ymm0, [eax]
     vmovdqu    ymm1, [eax + 32]
     lea        eax, [eax + 64]
     vpshufb    ymm0, ymm0, ymm5
     vpshufb    ymm1, ymm1, ymm5
-    sub        ecx, 16
     vmovdqu    [edx], ymm0
     vmovdqu    [edx + 32], ymm1
     lea        edx, [edx + 64]
+    sub        ecx, 16
     jg         wloop
 
     vzeroupper
@@ -6967,7 +5503,6 @@ void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
     jg         shuf_any1
     jmp        shuf99
 
-    align      4
   shuf_0123:
     movdqu     xmm0, [eax]
     lea        eax, [eax + 16]
@@ -6979,13 +5514,12 @@ void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
     pshufhw    xmm1, xmm1, 01Bh
     pshuflw    xmm1, xmm1, 01Bh
     packuswb   xmm0, xmm1
-    sub        ecx, 4
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 4
     jg         shuf_0123
     jmp        shuf99
 
-    align      4
   shuf_0321:
     movdqu     xmm0, [eax]
     lea        eax, [eax + 16]
@@ -6997,13 +5531,12 @@ void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
     pshufhw    xmm1, xmm1, 039h
     pshuflw    xmm1, xmm1, 039h
     packuswb   xmm0, xmm1
-    sub        ecx, 4
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 4
     jg         shuf_0321
     jmp        shuf99
 
-    align      4
   shuf_2103:
     movdqu     xmm0, [eax]
     lea        eax, [eax + 16]
@@ -7015,13 +5548,12 @@ void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
     pshufhw    xmm1, xmm1, 093h
     pshuflw    xmm1, xmm1, 093h
     packuswb   xmm0, xmm1
-    sub        ecx, 4
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 4
     jg         shuf_2103
     jmp        shuf99
 
-    align      4
   shuf_3012:
     movdqu     xmm0, [eax]
     lea        eax, [eax + 16]
@@ -7033,9 +5565,9 @@ void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
     pshufhw    xmm1, xmm1, 0C6h
     pshuflw    xmm1, xmm1, 0C6h
     packuswb   xmm0, xmm1
-    sub        ecx, 4
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 4
     jg         shuf_3012
 
   shuf99:
@@ -7066,7 +5598,6 @@ void I422ToYUY2Row_SSE2(const uint8* src_y,
     mov        ecx, [esp + 8 + 20]   // width
     sub        edx, esi
 
-    align      4
   convertloop:
     movq       xmm2, qword ptr [esi] // U
     movq       xmm3, qword ptr [esi + edx] // V
@@ -7104,7 +5635,6 @@ void I422ToUYVYRow_SSE2(const uint8* src_y,
     mov        ecx, [esp + 8 + 20]   // width
     sub        edx, esi
 
-    align      4
   convertloop:
     movq       xmm2, qword ptr [esi] // U
     movq       xmm3, qword ptr [esi + edx] // V
@@ -7141,7 +5671,6 @@ void ARGBPolynomialRow_SSE2(const uint8* src_argb,
     pxor       xmm3, xmm3  // 0 constant for zero extending bytes to ints.
 
     // 2 pixel loop.
-    align      4
  convertloop:
 //    pmovzxbd  xmm0, dword ptr [eax]  // BGRA pixel
 //    pmovzxbd  xmm4, dword ptr [eax + 4]  // BGRA pixel
@@ -7177,9 +5706,9 @@ void ARGBPolynomialRow_SSE2(const uint8* src_argb,
     cvttps2dq  xmm4, xmm4
     packuswb   xmm0, xmm4
     packuswb   xmm0, xmm0
-    sub        ecx, 2
     movq       qword ptr [edx], xmm0
     lea        edx, [edx + 8]
+    sub        ecx, 2
     jg         convertloop
     pop        esi
     ret
@@ -7203,7 +5732,6 @@ void ARGBPolynomialRow_AVX2(const uint8* src_argb,
     mov        ecx, [esp + 16]  /* width */
 
     // 2 pixel loop.
-    align      4
  convertloop:
     vpmovzxbd   ymm0, qword ptr [eax]  // 2 BGRA pixels
     lea         eax, [eax + 8]
@@ -7217,9 +5745,9 @@ void ARGBPolynomialRow_AVX2(const uint8* src_argb,
     vpackusdw   ymm0, ymm0, ymm0  // b0g0r0a0_00000000_b0g0r0a0_00000000
     vpermq      ymm0, ymm0, 0xd8  // b0g0r0a0_b0g0r0a0_00000000_00000000
     vpackuswb   xmm0, xmm0, xmm0  // bgrabgra_00000000_00000000_00000000
-    sub         ecx, 2
     vmovq       qword ptr [edx], xmm0
     lea         edx, [edx + 8]
+    sub         ecx, 2
     jg          convertloop
     vzeroupper
     ret
@@ -7239,7 +5767,6 @@ void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb,
     mov        ecx, [esp + 4 + 12]  /* width */
 
     // 1 pixel loop.
-    align      4
   convertloop:
     movzx      edx, byte ptr [eax]
     lea        eax, [eax + 4]
@@ -7273,7 +5800,6 @@ void RGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width) {
     mov        ecx, [esp + 4 + 12]  /* width */
 
     // 1 pixel loop.
-    align      4
   convertloop:
     movzx      edx, byte ptr [eax]
     lea        eax, [eax + 4]
@@ -7315,7 +5841,6 @@ void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
     pxor       xmm5, xmm5
 
     // 4 pixel loop.
-    align      4
   convertloop:
     movdqu     xmm0, qword ptr [eax]      // generate luma ptr
     pmaddubsw  xmm0, xmm3
@@ -7382,9 +5907,9 @@ void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
     movzx      edx, byte ptr [eax + 15]  // copy alpha.
     mov        byte ptr [edi + 15], dl
 
-    sub        ecx, 4
     lea        eax, [eax + 16]
     lea        edi, [edi + 16]
+    sub        ecx, 4
     jg         convertloop
 
     pop        edi
diff --git a/third_party/libyuv/source/scale.cc b/third_party/libyuv/source/scale.cc
index 5b33b5f04..482c5a61e 100644
--- a/third_party/libyuv/source/scale.cc
+++ b/third_party/libyuv/source/scale.cc
@@ -57,20 +57,15 @@ static void ScalePlaneDown2(int src_width, int src_height,
   if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 16)) {
     ScaleRowDown2 = filtering ? ScaleRowDown2Box_NEON : ScaleRowDown2_NEON;
   }
-#elif defined(HAS_SCALEROWDOWN2_SSE2)
+#endif
+#if defined(HAS_SCALEROWDOWN2_SSE2)
   if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 16)) {
-    ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_Unaligned_SSE2 :
-        (filtering == kFilterLinear ? ScaleRowDown2Linear_Unaligned_SSE2 :
-        ScaleRowDown2Box_Unaligned_SSE2);
-    if (IS_ALIGNED(src_ptr, 16) &&
-        IS_ALIGNED(src_stride, 16) && IS_ALIGNED(row_stride, 16) &&
-        IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
-      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_SSE2 :
-          (filtering == kFilterLinear ? ScaleRowDown2Linear_SSE2 :
-          ScaleRowDown2Box_SSE2);
-    }
+    ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_SSE2 :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_SSE2 :
+        ScaleRowDown2Box_SSE2);
   }
-#elif defined(HAS_SCALEROWDOWN2_MIPS_DSPR2)
+#endif
+#if defined(HAS_SCALEROWDOWN2_MIPS_DSPR2)
   if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(src_ptr, 4) &&
       IS_ALIGNED(src_stride, 4) && IS_ALIGNED(row_stride, 4) &&
       IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
@@ -112,21 +107,15 @@ static void ScalePlaneDown2_16(int src_width, int src_height,
     ScaleRowDown2 = filtering ? ScaleRowDown2Box_16_NEON :
         ScaleRowDown2_16_NEON;
   }
-#elif defined(HAS_SCALEROWDOWN2_16_SSE2)
+#endif
+#if defined(HAS_SCALEROWDOWN2_16_SSE2)
   if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 16)) {
-    ScaleRowDown2 = filtering == kFilterNone ?
-        ScaleRowDown2_Unaligned_16_SSE2 :
-        (filtering == kFilterLinear ? ScaleRowDown2Linear_Unaligned_16_SSE2 :
-        ScaleRowDown2Box_Unaligned_16_SSE2);
-    if (IS_ALIGNED(src_ptr, 16) &&
-        IS_ALIGNED(src_stride, 16) && IS_ALIGNED(row_stride, 16) &&
-        IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
-      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_16_SSE2 :
-          (filtering == kFilterLinear ? ScaleRowDown2Linear_16_SSE2 :
-          ScaleRowDown2Box_16_SSE2);
-    }
+    ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_16_SSE2 :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_16_SSE2 :
+        ScaleRowDown2Box_16_SSE2);
   }
-#elif defined(HAS_SCALEROWDOWN2_16_MIPS_DSPR2)
+#endif
+#if defined(HAS_SCALEROWDOWN2_16_MIPS_DSPR2)
   if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(src_ptr, 4) &&
       IS_ALIGNED(src_stride, 4) && IS_ALIGNED(row_stride, 4) &&
       IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
@@ -168,13 +157,13 @@ static void ScalePlaneDown4(int src_width, int src_height,
   if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 8)) {
     ScaleRowDown4 = filtering ? ScaleRowDown4Box_NEON : ScaleRowDown4_NEON;
   }
-#elif defined(HAS_SCALEROWDOWN4_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) &&
-      IS_ALIGNED(dst_width, 8) && IS_ALIGNED(row_stride, 16) &&
-      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+#endif
+#if defined(HAS_SCALEROWDOWN4_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
     ScaleRowDown4 = filtering ? ScaleRowDown4Box_SSE2 : ScaleRowDown4_SSE2;
   }
-#elif defined(HAS_SCALEROWDOWN4_MIPS_DSPR2)
+#endif
+#if defined(HAS_SCALEROWDOWN4_MIPS_DSPR2)
   if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(row_stride, 4) &&
       IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
       IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
@@ -212,14 +201,14 @@ static void ScalePlaneDown4_16(int src_width, int src_height,
     ScaleRowDown4 = filtering ? ScaleRowDown4Box_16_NEON :
         ScaleRowDown4_16_NEON;
   }
-#elif defined(HAS_SCALEROWDOWN4_16_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) &&
-      IS_ALIGNED(dst_width, 8) && IS_ALIGNED(row_stride, 16) &&
-      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+#endif
+#if defined(HAS_SCALEROWDOWN4_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
     ScaleRowDown4 = filtering ? ScaleRowDown4Box_16_SSE2 :
         ScaleRowDown4_16_SSE2;
   }
-#elif defined(HAS_SCALEROWDOWN4_16_MIPS_DSPR2)
+#endif
+#if defined(HAS_SCALEROWDOWN4_16_MIPS_DSPR2)
   if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(row_stride, 4) &&
       IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
       IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
@@ -271,8 +260,7 @@ static void ScalePlaneDown34(int src_width, int src_height,
   }
 #endif
 #if defined(HAS_SCALEROWDOWN34_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0) &&
-      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+  if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0)) {
     if (!filtering) {
       ScaleRowDown34_0 = ScaleRowDown34_SSSE3;
       ScaleRowDown34_1 = ScaleRowDown34_SSSE3;
@@ -351,8 +339,7 @@ static void ScalePlaneDown34_16(int src_width, int src_height,
   }
 #endif
 #if defined(HAS_SCALEROWDOWN34_16_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0) &&
-      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+  if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0)) {
     if (!filtering) {
       ScaleRowDown34_0 = ScaleRowDown34_16_SSSE3;
       ScaleRowDown34_1 = ScaleRowDown34_16_SSSE3;
@@ -445,9 +432,9 @@ static void ScalePlaneDown38(int src_width, int src_height,
       ScaleRowDown38_2 = ScaleRowDown38_2_Box_NEON;
     }
   }
-#elif defined(HAS_SCALEROWDOWN38_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0) &&
-      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+#endif
+#if defined(HAS_SCALEROWDOWN38_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0)) {
     if (!filtering) {
       ScaleRowDown38_3 = ScaleRowDown38_SSSE3;
       ScaleRowDown38_2 = ScaleRowDown38_SSSE3;
@@ -456,7 +443,8 @@ static void ScalePlaneDown38(int src_width, int src_height,
       ScaleRowDown38_2 = ScaleRowDown38_2_Box_SSSE3;
     }
   }
-#elif defined(HAS_SCALEROWDOWN38_MIPS_DSPR2)
+#endif
+#if defined(HAS_SCALEROWDOWN38_MIPS_DSPR2)
   if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 12 == 0) &&
       IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
       IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
@@ -522,9 +510,9 @@ static void ScalePlaneDown38_16(int src_width, int src_height,
       ScaleRowDown38_2 = ScaleRowDown38_2_Box_16_NEON;
     }
   }
-#elif defined(HAS_SCALEROWDOWN38_16_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0) &&
-      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+#endif
+#if defined(HAS_SCALEROWDOWN38_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0)) {
     if (!filtering) {
       ScaleRowDown38_3 = ScaleRowDown38_16_SSSE3;
       ScaleRowDown38_2 = ScaleRowDown38_16_SSSE3;
@@ -533,7 +521,8 @@ static void ScalePlaneDown38_16(int src_width, int src_height,
       ScaleRowDown38_2 = ScaleRowDown38_2_Box_16_SSSE3;
     }
   }
-#elif defined(HAS_SCALEROWDOWN38_16_MIPS_DSPR2)
+#endif
+#if defined(HAS_SCALEROWDOWN38_16_MIPS_DSPR2)
   if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 12 == 0) &&
       IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
       IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
@@ -758,11 +747,11 @@ static void ScalePlaneBox(int src_width, int src_height,
         uint16* dst_ptr, int src_width, int src_height) = ScaleAddRows_C;
 
 #if defined(HAS_SCALEADDROWS_SSE2)
-    if (TestCpuFlag(kCpuHasSSE2) &&
+    if (TestCpuFlag(kCpuHasSSE2)
 #ifdef AVOID_OVERREAD
-        IS_ALIGNED(src_width, 16) &&
+         && IS_ALIGNED(src_width, 16)
 #endif
-        IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+        ) {
       ScaleAddRows = ScaleAddRows_SSE2;
     }
 #endif
@@ -830,11 +819,11 @@ static void ScalePlaneBox_16(int src_width, int src_height,
         uint32* dst_ptr, int src_width, int src_height) = ScaleAddRows_16_C;
 
 #if defined(HAS_SCALEADDROWS_16_SSE2)
-    if (TestCpuFlag(kCpuHasSSE2) &&
+    if (TestCpuFlag(kCpuHasSSE2)
 #ifdef AVOID_OVERREAD
-        IS_ALIGNED(src_width, 16) &&
+        && IS_ALIGNED(src_width, 16)
 #endif
-        IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+        ) {
       ScaleAddRows = ScaleAddRows_16_SSE2;
     }
 #endif
@@ -886,29 +875,23 @@ void ScalePlaneBilinearDown(int src_width, int src_height,
   src_width = Abs(src_width);
 
 #if defined(HAS_INTERPOLATEROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && src_width >= 16) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     InterpolateRow = InterpolateRow_Any_SSE2;
     if (IS_ALIGNED(src_width, 16)) {
-      InterpolateRow = InterpolateRow_Unaligned_SSE2;
-      if (IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
-        InterpolateRow = InterpolateRow_SSE2;
-      }
+      InterpolateRow = InterpolateRow_SSE2;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && src_width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     InterpolateRow = InterpolateRow_Any_SSSE3;
     if (IS_ALIGNED(src_width, 16)) {
-      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
-        InterpolateRow = InterpolateRow_SSSE3;
-      }
+      InterpolateRow = InterpolateRow_SSSE3;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && src_width >= 32) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     InterpolateRow = InterpolateRow_Any_AVX2;
     if (IS_ALIGNED(src_width, 32)) {
       InterpolateRow = InterpolateRow_AVX2;
@@ -916,7 +899,7 @@ void ScalePlaneBilinearDown(int src_width, int src_height,
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && src_width >= 16) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     InterpolateRow = InterpolateRow_Any_NEON;
     if (IS_ALIGNED(src_width, 16)) {
       InterpolateRow = InterpolateRow_NEON;
@@ -924,7 +907,7 @@ void ScalePlaneBilinearDown(int src_width, int src_height,
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
-  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && src_width >= 4) {
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2)) {
     InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
     if (IS_ALIGNED(src_width, 4)) {
       InterpolateRow = InterpolateRow_MIPS_DSPR2;
@@ -988,29 +971,23 @@ void ScalePlaneBilinearDown_16(int src_width, int src_height,
   src_width = Abs(src_width);
 
 #if defined(HAS_INTERPOLATEROW_16_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && src_width >= 16) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     InterpolateRow = InterpolateRow_Any_16_SSE2;
     if (IS_ALIGNED(src_width, 16)) {
-      InterpolateRow = InterpolateRow_Unaligned_16_SSE2;
-      if (IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
-        InterpolateRow = InterpolateRow_16_SSE2;
-      }
+      InterpolateRow = InterpolateRow_16_SSE2;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_16_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && src_width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     InterpolateRow = InterpolateRow_Any_16_SSSE3;
     if (IS_ALIGNED(src_width, 16)) {
-      InterpolateRow = InterpolateRow_Unaligned_16_SSSE3;
-      if (IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
-        InterpolateRow = InterpolateRow_16_SSSE3;
-      }
+      InterpolateRow = InterpolateRow_16_SSSE3;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_16_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && src_width >= 32) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     InterpolateRow = InterpolateRow_Any_16_AVX2;
     if (IS_ALIGNED(src_width, 32)) {
       InterpolateRow = InterpolateRow_16_AVX2;
@@ -1018,7 +995,7 @@ void ScalePlaneBilinearDown_16(int src_width, int src_height,
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_16_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && src_width >= 16) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     InterpolateRow = InterpolateRow_Any_16_NEON;
     if (IS_ALIGNED(src_width, 16)) {
       InterpolateRow = InterpolateRow_16_NEON;
@@ -1026,7 +1003,7 @@ void ScalePlaneBilinearDown_16(int src_width, int src_height,
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_16_MIPS_DSPR2)
-  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && src_width >= 4) {
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2)) {
     InterpolateRow = InterpolateRow_Any_16_MIPS_DSPR2;
     if (IS_ALIGNED(src_width, 4)) {
       InterpolateRow = InterpolateRow_16_MIPS_DSPR2;
@@ -1087,29 +1064,23 @@ void ScalePlaneBilinearUp(int src_width, int src_height,
   src_width = Abs(src_width);
 
 #if defined(HAS_INTERPOLATEROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && dst_width >= 16) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     InterpolateRow = InterpolateRow_Any_SSE2;
     if (IS_ALIGNED(dst_width, 16)) {
-      InterpolateRow = InterpolateRow_Unaligned_SSE2;
-      if (IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
-        InterpolateRow = InterpolateRow_SSE2;
-      }
+      InterpolateRow = InterpolateRow_SSE2;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && dst_width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     InterpolateRow = InterpolateRow_Any_SSSE3;
     if (IS_ALIGNED(dst_width, 16)) {
-      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
-        InterpolateRow = InterpolateRow_SSSE3;
-      }
+      InterpolateRow = InterpolateRow_SSSE3;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && dst_width >= 32) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     InterpolateRow = InterpolateRow_Any_AVX2;
     if (IS_ALIGNED(dst_width, 32)) {
       InterpolateRow = InterpolateRow_AVX2;
@@ -1117,7 +1088,7 @@ void ScalePlaneBilinearUp(int src_width, int src_height,
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && dst_width >= 16) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     InterpolateRow = InterpolateRow_Any_NEON;
     if (IS_ALIGNED(dst_width, 16)) {
       InterpolateRow = InterpolateRow_NEON;
@@ -1125,7 +1096,7 @@ void ScalePlaneBilinearUp(int src_width, int src_height,
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
-  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && dst_width >= 4) {
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2)) {
     InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
     if (IS_ALIGNED(dst_width, 4)) {
       InterpolateRow = InterpolateRow_MIPS_DSPR2;
@@ -1144,9 +1115,7 @@ void ScalePlaneBilinearUp(int src_width, int src_height,
   if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
     ScaleFilterCols = ScaleColsUp2_C;
 #if defined(HAS_SCALECOLS_SSE2)
-    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
-        IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16) &&
-        IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
       ScaleFilterCols = ScaleColsUp2_SSE2;
     }
 #endif
@@ -1226,29 +1195,23 @@ void ScalePlaneBilinearUp_16(int src_width, int src_height,
   src_width = Abs(src_width);
 
 #if defined(HAS_INTERPOLATEROW_16_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && dst_width >= 16) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     InterpolateRow = InterpolateRow_Any_16_SSE2;
     if (IS_ALIGNED(dst_width, 16)) {
-      InterpolateRow = InterpolateRow_Unaligned_16_SSE2;
-      if (IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
-        InterpolateRow = InterpolateRow_16_SSE2;
-      }
+      InterpolateRow = InterpolateRow_16_SSE2;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_16_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && dst_width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     InterpolateRow = InterpolateRow_Any_16_SSSE3;
     if (IS_ALIGNED(dst_width, 16)) {
-      InterpolateRow = InterpolateRow_Unaligned_16_SSSE3;
-      if (IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
-        InterpolateRow = InterpolateRow_16_SSSE3;
-      }
+      InterpolateRow = InterpolateRow_16_SSSE3;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_16_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && dst_width >= 32) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     InterpolateRow = InterpolateRow_Any_16_AVX2;
     if (IS_ALIGNED(dst_width, 32)) {
       InterpolateRow = InterpolateRow_16_AVX2;
@@ -1256,7 +1219,7 @@ void ScalePlaneBilinearUp_16(int src_width, int src_height,
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_16_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && dst_width >= 16) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     InterpolateRow = InterpolateRow_Any_16_NEON;
     if (IS_ALIGNED(dst_width, 16)) {
       InterpolateRow = InterpolateRow_16_NEON;
@@ -1264,7 +1227,7 @@ void ScalePlaneBilinearUp_16(int src_width, int src_height,
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_16_MIPS_DSPR2)
-  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && dst_width >= 4) {
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2)) {
     InterpolateRow = InterpolateRow_Any_16_MIPS_DSPR2;
     if (IS_ALIGNED(dst_width, 4)) {
       InterpolateRow = InterpolateRow_16_MIPS_DSPR2;
@@ -1283,9 +1246,7 @@ void ScalePlaneBilinearUp_16(int src_width, int src_height,
   if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
     ScaleFilterCols = ScaleColsUp2_16_C;
 #if defined(HAS_SCALECOLS_16_SSE2)
-    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
-        IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16) &&
-        IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
       ScaleFilterCols = ScaleColsUp2_16_SSE2;
     }
 #endif
@@ -1366,9 +1327,7 @@ static void ScalePlaneSimple(int src_width, int src_height,
   if (src_width * 2 == dst_width && x < 0x8000) {
     ScaleCols = ScaleColsUp2_C;
 #if defined(HAS_SCALECOLS_SSE2)
-    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
-        IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16) &&
-        IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
       ScaleCols = ScaleColsUp2_SSE2;
     }
 #endif
@@ -1401,9 +1360,7 @@ static void ScalePlaneSimple_16(int src_width, int src_height,
   if (src_width * 2 == dst_width && x < 0x8000) {
     ScaleCols = ScaleColsUp2_16_C;
 #if defined(HAS_SCALECOLS_16_SSE2)
-    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
-        IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16) &&
-        IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
       ScaleCols = ScaleColsUp2_16_SSE2;
     }
 #endif
diff --git a/third_party/libyuv/source/scale_argb.cc b/third_party/libyuv/source/scale_argb.cc
index e339cd7c7..05b58e1ba 100644
--- a/third_party/libyuv/source/scale_argb.cc
+++ b/third_party/libyuv/source/scale_argb.cc
@@ -53,16 +53,14 @@ static void ScaleARGBDown2(int src_width, int src_height,
   }
 
 #if defined(HAS_SCALEARGBROWDOWN2_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 4) &&
-      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(row_stride, 16) &&
-      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 4)) {
     ScaleARGBRowDown2 = filtering == kFilterNone ? ScaleARGBRowDown2_SSE2 :
         (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_SSE2 :
         ScaleARGBRowDown2Box_SSE2);
   }
-#elif defined(HAS_SCALEARGBROWDOWN2_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 8) &&
-      IS_ALIGNED(src_argb, 4) && IS_ALIGNED(row_stride, 4)) {
+#endif
+#if defined(HAS_SCALEARGBROWDOWN2_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 8)) {
     ScaleARGBRowDown2 = filtering ? ScaleARGBRowDown2Box_NEON :
         ScaleARGBRowDown2_NEON;
   }
@@ -98,14 +96,12 @@ static void ScaleARGBDown4Box(int src_width, int src_height,
   assert(dx == 65536 * 4);  // Test scale factor of 4.
   assert((dy & 0x3ffff) == 0);  // Test vertical scale is multiple of 4.
 #if defined(HAS_SCALEARGBROWDOWN2_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 4) &&
-      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(row_stride, 16) &&
-      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 4)) {
     ScaleARGBRowDown2 = ScaleARGBRowDown2Box_SSE2;
   }
-#elif defined(HAS_SCALEARGBROWDOWN2_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 8) &&
-      IS_ALIGNED(src_argb, 4) && IS_ALIGNED(row_stride, 4)) {
+#endif
+#if defined(HAS_SCALEARGBROWDOWN2_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 8)) {
     ScaleARGBRowDown2 = ScaleARGBRowDown2Box_NEON;
   }
 #endif
@@ -139,14 +135,13 @@ static void ScaleARGBDownEven(int src_width, int src_height,
   assert(IS_ALIGNED(src_height, 2));
   src_argb += (y >> 16) * src_stride + (x >> 16) * 4;
 #if defined(HAS_SCALEARGBROWDOWNEVEN_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 4) &&
-      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 4)) {
     ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_SSE2 :
         ScaleARGBRowDownEven_SSE2;
   }
-#elif defined(HAS_SCALEARGBROWDOWNEVEN_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 4) &&
-      IS_ALIGNED(src_argb, 4)) {
+#endif
+#if defined(HAS_SCALEARGBROWDOWNEVEN_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 4)) {
     ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_NEON :
         ScaleARGBRowDownEven_NEON;
   }
@@ -190,29 +185,23 @@ static void ScaleARGBBilinearDown(int src_width, int src_height,
   src_argb += xl * 4;
   x -= (int)(xl << 16);
 #if defined(HAS_INTERPOLATEROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && clip_src_width >= 16) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     InterpolateRow = InterpolateRow_Any_SSE2;
     if (IS_ALIGNED(clip_src_width, 16)) {
-      InterpolateRow = InterpolateRow_Unaligned_SSE2;
-      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16)) {
-        InterpolateRow = InterpolateRow_SSE2;
-      }
+      InterpolateRow = InterpolateRow_SSE2;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && clip_src_width >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     InterpolateRow = InterpolateRow_Any_SSSE3;
     if (IS_ALIGNED(clip_src_width, 16)) {
-      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16)) {
-        InterpolateRow = InterpolateRow_SSSE3;
-      }
+      InterpolateRow = InterpolateRow_SSSE3;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && clip_src_width >= 32) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     InterpolateRow = InterpolateRow_Any_AVX2;
     if (IS_ALIGNED(clip_src_width, 32)) {
       InterpolateRow = InterpolateRow_AVX2;
@@ -220,15 +209,15 @@ static void ScaleARGBBilinearDown(int src_width, int src_height,
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && clip_src_width >= 16) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     InterpolateRow = InterpolateRow_Any_NEON;
     if (IS_ALIGNED(clip_src_width, 16)) {
       InterpolateRow = InterpolateRow_NEON;
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROWS_MIPS_DSPR2)
-  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && clip_src_width >= 4 &&
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
       IS_ALIGNED(src_argb, 4) && IS_ALIGNED(src_stride, 4)) {
     InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
     if (IS_ALIGNED(clip_src_width, 4)) {
@@ -286,29 +275,23 @@ static void ScaleARGBBilinearUp(int src_width, int src_height,
       filtering ? ScaleARGBFilterCols_C : ScaleARGBCols_C;
   const int max_y = (src_height - 1) << 16;
 #if defined(HAS_INTERPOLATEROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && dst_width >= 4) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     InterpolateRow = InterpolateRow_Any_SSE2;
     if (IS_ALIGNED(dst_width, 4)) {
-      InterpolateRow = InterpolateRow_Unaligned_SSE2;
-      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
-        InterpolateRow = InterpolateRow_SSE2;
-      }
+      InterpolateRow = InterpolateRow_SSE2;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && dst_width >= 4) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     InterpolateRow = InterpolateRow_Any_SSSE3;
     if (IS_ALIGNED(dst_width, 4)) {
-      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
-        InterpolateRow = InterpolateRow_SSSE3;
-      }
+      InterpolateRow = InterpolateRow_SSSE3;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && dst_width >= 8) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     InterpolateRow = InterpolateRow_Any_AVX2;
     if (IS_ALIGNED(dst_width, 8)) {
       InterpolateRow = InterpolateRow_AVX2;
@@ -316,15 +299,15 @@ static void ScaleARGBBilinearUp(int src_width, int src_height,
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && dst_width >= 4) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     InterpolateRow = InterpolateRow_Any_NEON;
     if (IS_ALIGNED(dst_width, 4)) {
       InterpolateRow = InterpolateRow_NEON;
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROWS_MIPS_DSPR2)
-  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && dst_width >= 1 &&
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
       IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride, 4)) {
     InterpolateRow = InterpolateRow_MIPS_DSPR2;
   }
@@ -346,9 +329,7 @@ static void ScaleARGBBilinearUp(int src_width, int src_height,
   if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
     ScaleARGBFilterCols = ScaleARGBColsUp2_C;
 #if defined(HAS_SCALEARGBCOLSUP2_SSE2)
-    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
-        IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
-        IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
       ScaleARGBFilterCols = ScaleARGBColsUp2_SSE2;
     }
 #endif
@@ -427,18 +408,15 @@ static void ScaleYUVToARGBBilinearUp(int src_width, int src_height,
                         uint8* rgb_buf,
                         int width) = I422ToARGBRow_C;
 #if defined(HAS_I422TOARGBROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && src_width >= 8) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
     if (IS_ALIGNED(src_width, 8)) {
-      I422ToARGBRow = I422ToARGBRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
-        I422ToARGBRow = I422ToARGBRow_SSSE3;
-      }
+      I422ToARGBRow = I422ToARGBRow_SSSE3;
     }
   }
 #endif
 #if defined(HAS_I422TOARGBROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && src_width >= 16) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     I422ToARGBRow = I422ToARGBRow_Any_AVX2;
     if (IS_ALIGNED(src_width, 16)) {
       I422ToARGBRow = I422ToARGBRow_AVX2;
@@ -446,7 +424,7 @@ static void ScaleYUVToARGBBilinearUp(int src_width, int src_height,
   }
 #endif
 #if defined(HAS_I422TOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && src_width >= 8) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     I422ToARGBRow = I422ToARGBRow_Any_NEON;
     if (IS_ALIGNED(src_width, 8)) {
       I422ToARGBRow = I422ToARGBRow_NEON;
@@ -467,29 +445,23 @@ static void ScaleYUVToARGBBilinearUp(int src_width, int src_height,
       ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
       InterpolateRow_C;
 #if defined(HAS_INTERPOLATEROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && dst_width >= 4) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     InterpolateRow = InterpolateRow_Any_SSE2;
     if (IS_ALIGNED(dst_width, 4)) {
-      InterpolateRow = InterpolateRow_Unaligned_SSE2;
-      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
-        InterpolateRow = InterpolateRow_SSE2;
-      }
+      InterpolateRow = InterpolateRow_SSE2;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && dst_width >= 4) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     InterpolateRow = InterpolateRow_Any_SSSE3;
     if (IS_ALIGNED(dst_width, 4)) {
-      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
-        InterpolateRow = InterpolateRow_SSSE3;
-      }
+      InterpolateRow = InterpolateRow_SSSE3;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && dst_width >= 8) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     InterpolateRow = InterpolateRow_Any_AVX2;
     if (IS_ALIGNED(dst_width, 8)) {
       InterpolateRow = InterpolateRow_AVX2;
@@ -497,15 +469,15 @@ static void ScaleYUVToARGBBilinearUp(int src_width, int src_height,
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && dst_width >= 4) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     InterpolateRow = InterpolateRow_Any_NEON;
     if (IS_ALIGNED(dst_width, 4)) {
       InterpolateRow = InterpolateRow_NEON;
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROWS_MIPS_DSPR2)
-  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && dst_width >= 1 &&
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
       IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
     InterpolateRow = InterpolateRow_MIPS_DSPR2;
   }
@@ -531,9 +503,7 @@ static void ScaleYUVToARGBBilinearUp(int src_width, int src_height,
   if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
     ScaleARGBFilterCols = ScaleARGBColsUp2_C;
 #if defined(HAS_SCALEARGBCOLSUP2_SSE2)
-    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
-        IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
-        IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
       ScaleARGBFilterCols = ScaleARGBColsUp2_SSE2;
     }
 #endif
@@ -640,9 +610,7 @@ static void ScaleARGBSimple(int src_width, int src_height,
   if (src_width * 2 == dst_width && x < 0x8000) {
     ScaleARGBCols = ScaleARGBColsUp2_C;
 #if defined(HAS_SCALEARGBCOLSUP2_SSE2)
-    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
-        IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
-        IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
       ScaleARGBCols = ScaleARGBColsUp2_SSE2;
     }
 #endif
diff --git a/third_party/libyuv/source/scale_common.cc b/third_party/libyuv/source/scale_common.cc
index e4b2acc41..96e2564b0 100644
--- a/third_party/libyuv/source/scale_common.cc
+++ b/third_party/libyuv/source/scale_common.cc
@@ -885,31 +885,23 @@ void ScalePlaneVertical(int src_height,
   assert(dst_height > 0);
   src_argb += (x >> 16) * bpp;
 #if defined(HAS_INTERPOLATEROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && dst_width_bytes >= 16) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     InterpolateRow = InterpolateRow_Any_SSE2;
     if (IS_ALIGNED(dst_width_bytes, 16)) {
-      InterpolateRow = InterpolateRow_Unaligned_SSE2;
-      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
-          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
-        InterpolateRow = InterpolateRow_SSE2;
-      }
+      InterpolateRow = InterpolateRow_SSE2;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && dst_width_bytes >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     InterpolateRow = InterpolateRow_Any_SSSE3;
     if (IS_ALIGNED(dst_width_bytes, 16)) {
-      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
-      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
-          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
-        InterpolateRow = InterpolateRow_SSSE3;
-      }
+      InterpolateRow = InterpolateRow_SSSE3;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && dst_width_bytes >= 32) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     InterpolateRow = InterpolateRow_Any_AVX2;
     if (IS_ALIGNED(dst_width_bytes, 32)) {
       InterpolateRow = InterpolateRow_AVX2;
@@ -917,15 +909,15 @@ void ScalePlaneVertical(int src_height,
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && dst_width_bytes >= 16) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     InterpolateRow = InterpolateRow_Any_NEON;
     if (IS_ALIGNED(dst_width_bytes, 16)) {
       InterpolateRow = InterpolateRow_NEON;
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROWS_MIPS_DSPR2)
-  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && dst_width_bytes >= 4 &&
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
       IS_ALIGNED(src_argb, 4) && IS_ALIGNED(src_stride, 4) &&
       IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride, 4)) {
     InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
@@ -967,31 +959,23 @@ void ScalePlaneVertical_16(int src_height,
   assert(dst_height > 0);
   src_argb += (x >> 16) * wpp;
 #if defined(HAS_INTERPOLATEROW_16_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && dst_width_bytes >= 16) {
+  if (TestCpuFlag(kCpuHasSSE2)) {
     InterpolateRow = InterpolateRow_Any_16_SSE2;
     if (IS_ALIGNED(dst_width_bytes, 16)) {
-      InterpolateRow = InterpolateRow_Unaligned_16_SSE2;
-      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
-          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
-        InterpolateRow = InterpolateRow_16_SSE2;
-      }
+      InterpolateRow = InterpolateRow_16_SSE2;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_16_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && dst_width_bytes >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     InterpolateRow = InterpolateRow_Any_16_SSSE3;
     if (IS_ALIGNED(dst_width_bytes, 16)) {
-      InterpolateRow = InterpolateRow_Unaligned_16_SSSE3;
-      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
-          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
-        InterpolateRow = InterpolateRow_16_SSSE3;
-      }
+      InterpolateRow = InterpolateRow_16_SSSE3;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_16_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && dst_width_bytes >= 32) {
+  if (TestCpuFlag(kCpuHasAVX2)) {
     InterpolateRow = InterpolateRow_Any_16_AVX2;
     if (IS_ALIGNED(dst_width_bytes, 32)) {
       InterpolateRow = InterpolateRow_16_AVX2;
@@ -999,15 +983,15 @@ void ScalePlaneVertical_16(int src_height,
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_16_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && dst_width_bytes >= 16) {
+  if (TestCpuFlag(kCpuHasNEON)) {
     InterpolateRow = InterpolateRow_Any_16_NEON;
     if (IS_ALIGNED(dst_width_bytes, 16)) {
       InterpolateRow = InterpolateRow_16_NEON;
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROWS_16_MIPS_DSPR2)
-  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && dst_width_bytes >= 4 &&
+#if defined(HAS_INTERPOLATEROW_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
       IS_ALIGNED(src_argb, 4) && IS_ALIGNED(src_stride, 4) &&
       IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride, 4)) {
     InterpolateRow = InterpolateRow_Any_16_MIPS_DSPR2;
diff --git a/third_party/libyuv/source/scale_neon.cc b/third_party/libyuv/source/scale_neon.cc
index 1b8a5ba58..7921219b5 100644
--- a/third_party/libyuv/source/scale_neon.cc
+++ b/third_party/libyuv/source/scale_neon.cc
@@ -16,7 +16,8 @@ extern "C" {
 #endif
 
 // This module is for GCC Neon.
-#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__)
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \
+    !defined(__aarch64__)
 
 // NEON downscalers with interpolation.
 // Provided by Fritz Koenig
@@ -756,7 +757,7 @@ void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, ptrdiff_t src_stride,
   );
 }
 
-#endif  // __ARM_NEON__
+#endif  // defined(__ARM_NEON__) && !defined(__aarch64__)
 
 #ifdef __cplusplus
 }  // extern "C"
diff --git a/third_party/libyuv/source/scale_neon64.cc b/third_party/libyuv/source/scale_neon64.cc
index 64c7d10db..fb68b67d2 100644
--- a/third_party/libyuv/source/scale_neon64.cc
+++ b/third_party/libyuv/source/scale_neon64.cc
@@ -8,133 +8,122 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
+#include "libyuv/scale.h"
 #include "libyuv/row.h"
+#include "libyuv/scale_row.h"
 
 #ifdef __cplusplus
 namespace libyuv {
 extern "C" {
 #endif
 
-// This module is for GCC Neon.
+// This module is for GCC Neon armv8 64 bit.
 #if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
-#ifdef HAS_SCALEROWDOWN2_NEON
+
 // Read 32x1 throw away even pixels, and write 16x1.
 void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
                         uint8* dst, int dst_width) {
   asm volatile (
-    ".p2align   2                              \n"
   "1:                                          \n"
-    // load even pixels into q0, odd into q1
+    // load even pixels into v0, odd into v1
     MEMACCESS(0)
-    "vld2.8     {q0, q1}, [%0]!                \n"
+    "ld2        {v0.16b,v1.16b}, [%0], #32    \n"
     "subs       %2, %2, #16                    \n"  // 16 processed per loop
     MEMACCESS(1)
-    "vst1.8     {q1}, [%1]!                    \n"  // store odd pixels
-    "bgt        1b                             \n"
+    "st1        {v1.16b}, [%1], #16            \n"  // store odd pixels
+    "b.gt       1b                             \n"
   : "+r"(src_ptr),          // %0
     "+r"(dst),              // %1
     "+r"(dst_width)         // %2
   :
-  : "q0", "q1"              // Clobber List
+  : "v0", "v1"              // Clobber List
   );
 }
-#endif //HAS_SCALEROWDOWN2_NEON
 
-#ifdef HAS_SCALEROWDOWN2_NEON
 // Read 32x2 average down and write 16x1.
 void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
                            uint8* dst, int dst_width) {
   asm volatile (
     // change the stride to row 2 pointer
-    "add        %1, %0                         \n"
-    ".p2align   2                              \n"
+    "add        %1, %1, %0                     \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "vld1.8     {q0, q1}, [%0]!                \n"  // load row 1 and post inc
+    "ld1        {v0.16b,v1.16b}, [%0], #32    \n"  // load row 1 and post inc
     MEMACCESS(1)
-    "vld1.8     {q2, q3}, [%1]!                \n"  // load row 2 and post inc
+    "ld1        {v2.16b, v3.16b}, [%1], #32    \n"  // load row 2 and post inc
     "subs       %3, %3, #16                    \n"  // 16 processed per loop
-    "vpaddl.u8  q0, q0                         \n"  // row 1 add adjacent
-    "vpaddl.u8  q1, q1                         \n"
-    "vpadal.u8  q0, q2                         \n"  // row 2 add adjacent + row1
-    "vpadal.u8  q1, q3                         \n"
-    "vrshrn.u16 d0, q0, #2                     \n"  // downshift, round and pack
-    "vrshrn.u16 d1, q1, #2                     \n"
+    "uaddlp     v0.8h, v0.16b                  \n"  // row 1 add adjacent
+    "uaddlp     v1.8h, v1.16b                  \n"
+    "uadalp     v0.8h, v2.16b                  \n"  // row 2 add adjacent + row1
+    "uadalp     v1.8h, v3.16b                  \n"
+    "rshrn      v0.8b, v0.8h, #2               \n"  // downshift, round and pack
+    "rshrn2     v0.16b, v1.8h, #2              \n"
     MEMACCESS(2)
-    "vst1.8     {q0}, [%2]!                    \n"
-    "bgt        1b                             \n"
+    "st1        {v0.16b}, [%2], #16            \n"
+    "b.gt       1b                             \n"
   : "+r"(src_ptr),          // %0
     "+r"(src_stride),       // %1
     "+r"(dst),              // %2
     "+r"(dst_width)         // %3
   :
-  : "q0", "q1", "q2", "q3"     // Clobber List
+  : "v0", "v1", "v2", "v3"     // Clobber List
   );
 }
-#endif //HAS_SCALEROWDOWN2_NEON
 
-#ifdef HAS_SCALEROWDOWN4_NEON
 void ScaleRowDown4_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
                         uint8* dst_ptr, int dst_width) {
   asm volatile (
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n" // src line 0
-    "subs       %2, %2, #8                     \n" // 8 processed per loop
+    "ld4     {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32          \n"  // src line 0
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop
     MEMACCESS(1)
-    "vst1.8     {d2}, [%1]!                    \n"
-    "bgt        1b                             \n"
+    "st1     {v2.8b}, [%1], #8                 \n"
+    "b.gt       1b                             \n"
   : "+r"(src_ptr),          // %0
     "+r"(dst_ptr),          // %1
     "+r"(dst_width)         // %2
   :
-  : "q0", "q1", "memory", "cc"
+  : "v0", "v1", "v2", "v3", "memory", "cc"
   );
 }
-#endif //HAS_SCALEROWDOWN4_NEON
 
-#ifdef HAS_SCALEROWDOWN4_NEON
 void ScaleRowDown4Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
                            uint8* dst_ptr, int dst_width) {
   const uint8* src_ptr1 = src_ptr + src_stride;
   const uint8* src_ptr2 = src_ptr + src_stride * 2;
   const uint8* src_ptr3 = src_ptr + src_stride * 3;
 asm volatile (
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"   // load up 16x4
+    "ld1     {v0.16b}, [%0], #16               \n"   // load up 16x4
     MEMACCESS(3)
-    "vld1.8     {q1}, [%3]!                    \n"
+    "ld1     {v1.16b}, [%2], #16               \n"
     MEMACCESS(4)
-    "vld1.8     {q2}, [%4]!                    \n"
+    "ld1     {v2.16b}, [%3], #16               \n"
     MEMACCESS(5)
-    "vld1.8     {q3}, [%5]!                    \n"
-    "subs       %2, %2, #4                     \n"
-    "vpaddl.u8  q0, q0                         \n"
-    "vpadal.u8  q0, q1                         \n"
-    "vpadal.u8  q0, q2                         \n"
-    "vpadal.u8  q0, q3                         \n"
-    "vpaddl.u16 q0, q0                         \n"
-    "vrshrn.u32 d0, q0, #4                     \n"   // divide by 16 w/rounding
-    "vmovn.u16  d0, q0                         \n"
+    "ld1     {v3.16b}, [%4], #16               \n"
+    "subs    %5, %5, #4                        \n"
+    "uaddlp  v0.8h, v0.16b                     \n"
+    "uadalp  v0.8h, v1.16b                     \n"
+    "uadalp  v0.8h, v2.16b                     \n"
+    "uadalp  v0.8h, v3.16b                     \n"
+    "addp    v0.8h, v0.8h, v0.8h               \n"
+    "rshrn   v0.8b, v0.8h, #4                  \n"   // divide by 16 w/rounding
     MEMACCESS(1)
-    "vst1.32    {d0[0]}, [%1]!                 \n"
-    "bgt        1b                             \n"
+    "st1    {v0.s}[0], [%1], #4                \n"
+    "b.gt       1b                             \n"
   : "+r"(src_ptr),   // %0
     "+r"(dst_ptr),   // %1
-    "+r"(dst_width), // %2
-    "+r"(src_ptr1),  // %3
-    "+r"(src_ptr2),  // %4
-    "+r"(src_ptr3)   // %5
+    "+r"(src_ptr1),  // %2
+    "+r"(src_ptr2),  // %3
+    "+r"(src_ptr3),  // %4
+    "+r"(dst_width)  // %5
   :
-  : "q0", "q1", "q2", "q3", "memory", "cc"
+  : "v0", "v1", "v2", "v3", "memory", "cc"
   );
 }
-#endif //HAS_SCALEROWDOWN4_NEON
 
-#ifdef HAS_SCALEROWDOWN34_NEON
 // Down scale from 4 to 3 pixels. Use the neon multilane read/write
 // to load up the every 4th pixel into a 4 different registers.
 // Point samples 32 pixels to 24 pixels.
@@ -142,136 +131,129 @@ void ScaleRowDown34_NEON(const uint8* src_ptr,
                          ptrdiff_t src_stride,
                          uint8* dst_ptr, int dst_width) {
   asm volatile (
-    ".p2align   2                              \n"
-  "1:                                          \n"
+  "1:                                                  \n"
     MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!      \n" // src line 0
-    "subs       %2, %2, #24                  \n"
-    "vmov       d2, d3                       \n" // order d0, d1, d2
+    "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32                \n"  // src line 0
+    "subs      %2, %2, #24                             \n"
+    "orr       v2.16b, v3.16b, v3.16b                  \n"  // order v0, v1, v2
     MEMACCESS(1)
-    "vst3.8     {d0, d1, d2}, [%1]!          \n"
-    "bgt        1b                           \n"
+    "st3       {v0.8b,v1.8b,v2.8b}, [%1], #24                \n"
+    "b.gt      1b                                      \n"
   : "+r"(src_ptr),          // %0
     "+r"(dst_ptr),          // %1
     "+r"(dst_width)         // %2
   :
-  : "d0", "d1", "d2", "d3", "memory", "cc"
+  : "v0", "v1", "v2", "v3", "memory", "cc"
   );
 }
-#endif //HAS_SCALEROWDOWN34_NEON
 
-#ifdef HAS_SCALEROWDOWN34_NEON
 void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr,
                                ptrdiff_t src_stride,
                                uint8* dst_ptr, int dst_width) {
   asm volatile (
-    "vmov.u8    d24, #3                        \n"
-    "add        %3, %0                         \n"
-    ".p2align   2                              \n"
-  "1:                                          \n"
+    "movi      v20.8b, #3                              \n"
+    "add       %3, %3, %0                              \n"
+  "1:                                                  \n"
     MEMACCESS(0)
-    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n" // src line 0
+    "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32                \n"  // src line 0
     MEMACCESS(3)
-    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n" // src line 1
-    "subs         %2, %2, #24                  \n"
+    "ld4       {v4.8b,v5.8b,v6.8b,v7.8b}, [%3], #32                \n"  // src line 1
+    "subs         %2, %2, #24                          \n"
 
     // filter src line 0 with src line 1
     // expand chars to shorts to allow for room
     // when adding lines together
-    "vmovl.u8     q8, d4                       \n"
-    "vmovl.u8     q9, d5                       \n"
-    "vmovl.u8     q10, d6                      \n"
-    "vmovl.u8     q11, d7                      \n"
+    "ushll     v16.8h, v4.8b, #0                       \n"
+    "ushll     v17.8h, v5.8b, #0                       \n"
+    "ushll     v18.8h, v6.8b, #0                       \n"
+    "ushll     v19.8h, v7.8b, #0                       \n"
 
     // 3 * line_0 + line_1
-    "vmlal.u8     q8, d0, d24                  \n"
-    "vmlal.u8     q9, d1, d24                  \n"
-    "vmlal.u8     q10, d2, d24                 \n"
-    "vmlal.u8     q11, d3, d24                 \n"
+    "umlal     v16.8h, v0.8b, v20.8b                   \n"
+    "umlal     v17.8h, v1.8b, v20.8b                   \n"
+    "umlal     v18.8h, v2.8b, v20.8b                   \n"
+    "umlal     v19.8h, v3.8b, v20.8b                   \n"
 
     // (3 * line_0 + line_1) >> 2
-    "vqrshrn.u16  d0, q8, #2                   \n"
-    "vqrshrn.u16  d1, q9, #2                   \n"
-    "vqrshrn.u16  d2, q10, #2                  \n"
-    "vqrshrn.u16  d3, q11, #2                  \n"
+    "uqrshrn   v0.8b, v16.8h, #2                       \n"
+    "uqrshrn   v1.8b, v17.8h, #2                       \n"
+    "uqrshrn   v2.8b, v18.8h, #2                       \n"
+    "uqrshrn   v3.8b, v19.8h, #2                       \n"
 
     // a0 = (src[0] * 3 + s[1] * 1) >> 2
-    "vmovl.u8     q8, d1                       \n"
-    "vmlal.u8     q8, d0, d24                  \n"
-    "vqrshrn.u16  d0, q8, #2                   \n"
+    "ushll     v16.8h, v1.8b, #0                       \n"
+    "umlal     v16.8h, v0.8b, v20.8b                   \n"
+    "uqrshrn   v0.8b, v16.8h, #2                       \n"
 
     // a1 = (src[1] * 1 + s[2] * 1) >> 1
-    "vrhadd.u8    d1, d1, d2                   \n"
+    "urhadd    v1.8b, v1.8b, v2.8b                     \n"
 
     // a2 = (src[2] * 1 + s[3] * 3) >> 2
-    "vmovl.u8     q8, d2                       \n"
-    "vmlal.u8     q8, d3, d24                  \n"
-    "vqrshrn.u16  d2, q8, #2                   \n"
+    "ushll     v16.8h, v2.8b, #0                       \n"
+    "umlal     v16.8h, v3.8b, v20.8b                   \n"
+    "uqrshrn   v2.8b, v16.8h, #2                       \n"
 
     MEMACCESS(1)
-    "vst3.8       {d0, d1, d2}, [%1]!          \n"
+    "st3       {v0.8b,v1.8b,v2.8b}, [%1], #24                \n"
 
-    "bgt          1b                           \n"
+    "b.gt      1b                                      \n"
   : "+r"(src_ptr),          // %0
     "+r"(dst_ptr),          // %1
     "+r"(dst_width),        // %2
     "+r"(src_stride)        // %3
   :
-  : "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11", "d24", "memory", "cc"
+  : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19",
+    "v20", "memory", "cc"
   );
 }
-#endif //ScaleRowDown34_0_Box_NEON
 
-#ifdef HAS_SCALEROWDOWN34_NEON
 void ScaleRowDown34_1_Box_NEON(const uint8* src_ptr,
                                ptrdiff_t src_stride,
                                uint8* dst_ptr, int dst_width) {
   asm volatile (
-    "vmov.u8    d24, #3                        \n"
-    "add        %3, %0                         \n"
-    ".p2align   2                              \n"
-  "1:                                          \n"
+    "movi      v20.8b, #3                              \n"
+    "add       %3, %3, %0                              \n"
+  "1:                                                  \n"
     MEMACCESS(0)
-    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n" // src line 0
+    "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32                \n"  // src line 0
     MEMACCESS(3)
-    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n" // src line 1
-    "subs         %2, %2, #24                  \n"
+    "ld4       {v4.8b,v5.8b,v6.8b,v7.8b}, [%3], #32                \n"  // src line 1
+    "subs         %2, %2, #24                          \n"
     // average src line 0 with src line 1
-    "vrhadd.u8    q0, q0, q2                   \n"
-    "vrhadd.u8    q1, q1, q3                   \n"
+    "urhadd    v0.8b, v0.8b, v4.8b                     \n"
+    "urhadd    v1.8b, v1.8b, v5.8b                     \n"
+    "urhadd    v2.8b, v2.8b, v6.8b                     \n"
+    "urhadd    v3.8b, v3.8b, v7.8b                     \n"
 
     // a0 = (src[0] * 3 + s[1] * 1) >> 2
-    "vmovl.u8     q3, d1                       \n"
-    "vmlal.u8     q3, d0, d24                  \n"
-    "vqrshrn.u16  d0, q3, #2                   \n"
+    "ushll     v4.8h, v1.8b, #0                        \n"
+    "umlal     v4.8h, v0.8b, v20.8b                    \n"
+    "uqrshrn   v0.8b, v4.8h, #2                        \n"
 
     // a1 = (src[1] * 1 + s[2] * 1) >> 1
-    "vrhadd.u8    d1, d1, d2                   \n"
+    "urhadd    v1.8b, v1.8b, v2.8b                     \n"
 
     // a2 = (src[2] * 1 + s[3] * 3) >> 2
-    "vmovl.u8     q3, d2                       \n"
-    "vmlal.u8     q3, d3, d24                  \n"
-    "vqrshrn.u16  d2, q3, #2                   \n"
+    "ushll     v4.8h, v2.8b, #0                        \n"
+    "umlal     v4.8h, v3.8b, v20.8b                    \n"
+    "uqrshrn   v2.8b, v4.8h, #2                        \n"
 
     MEMACCESS(1)
-    "vst3.8       {d0, d1, d2}, [%1]!          \n"
-    "bgt          1b                           \n"
+    "st3       {v0.8b,v1.8b,v2.8b}, [%1], #24                \n"
+    "b.gt      1b                                      \n"
   : "+r"(src_ptr),          // %0
     "+r"(dst_ptr),          // %1
     "+r"(dst_width),        // %2
     "+r"(src_stride)        // %3
   :
-  : "r4", "q0", "q1", "q2", "q3", "d24", "memory", "cc"
+  : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", "memory", "cc"
   );
 }
-#endif //HAS_SCALEROWDOWN34_NEON
 
-#ifdef HAS_SCALEROWDOWN38_NEON
-#define HAS_SCALEROWDOWN38_NEON
 static uvec8 kShuf38 =
   { 0, 3, 6, 8, 11, 14, 16, 19, 22, 24, 27, 30, 0, 0, 0, 0 };
 static uvec8 kShuf38_2 =
-  { 0, 8, 16, 2, 10, 17, 4, 12, 18, 6, 14, 19, 0, 0, 0, 0 };
+  { 0, 16, 32, 2, 18, 33, 4, 20, 34, 6, 22, 35, 0, 0, 0, 0 };
 static vec16 kMult38_Div6 =
   { 65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12,
     65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12 };
@@ -285,504 +267,498 @@ void ScaleRowDown38_NEON(const uint8* src_ptr,
                          uint8* dst_ptr, int dst_width) {
   asm volatile (
     MEMACCESS(3)
-    "vld1.8     {q3}, [%3]                     \n"
-    ".p2align   2                              \n"
-  "1:                                          \n"
+    "ld1       {v3.16b}, [%3]                          \n"
+  "1:                                                  \n"
     MEMACCESS(0)
-    "vld1.8     {d0, d1, d2, d3}, [%0]!        \n"
-    "subs       %2, %2, #12                    \n"
-    "vtbl.u8    d4, {d0, d1, d2, d3}, d6       \n"
-    "vtbl.u8    d5, {d0, d1, d2, d3}, d7       \n"
+    "ld1       {v0.16b,v1.16b}, [%0], #32             \n"
+    "subs      %2, %2, #12                             \n"
+    "tbl       v2.16b, {v0.16b,v1.16b}, v3.16b        \n"
     MEMACCESS(1)
-    "vst1.8     {d4}, [%1]!                    \n"
+    "st1       {v2.8b}, [%1], #8                       \n"
     MEMACCESS(1)
-    "vst1.32    {d5[0]}, [%1]!                 \n"
-    "bgt        1b                             \n"
+    "st1       {v2.s}[2], [%1], #4                     \n"
+    "b.gt      1b                                      \n"
   : "+r"(src_ptr),          // %0
     "+r"(dst_ptr),          // %1
     "+r"(dst_width)         // %2
   : "r"(&kShuf38)           // %3
-  : "d0", "d1", "d2", "d3", "d4", "d5", "memory", "cc"
+  : "v0", "v1", "v2", "v3", "memory", "cc"
   );
 }
 
-#endif //HAS_SCALEROWDOWN38_NEON
-
-#ifdef HAS_SCALEROWDOWN38_NEON
 // 32x3 -> 12x1
 void OMITFP ScaleRowDown38_3_Box_NEON(const uint8* src_ptr,
                                       ptrdiff_t src_stride,
                                       uint8* dst_ptr, int dst_width) {
   const uint8* src_ptr1 = src_ptr + src_stride * 2;
+  ptrdiff_t tmp_src_stride = src_stride;
 
   asm volatile (
     MEMACCESS(5)
-    "vld1.16    {q13}, [%5]                    \n"
+    "ld1       {v29.8h}, [%5]                          \n"
     MEMACCESS(6)
-    "vld1.8     {q14}, [%6]                    \n"
+    "ld1       {v30.16b}, [%6]                         \n"
     MEMACCESS(7)
-    "vld1.8     {q15}, [%7]                    \n"
-    "add        %3, %0                         \n"
-    ".p2align   2                              \n"
-  "1:                                          \n"
-
-    // d0 = 00 40 01 41 02 42 03 43
-    // d1 = 10 50 11 51 12 52 13 53
-    // d2 = 20 60 21 61 22 62 23 63
-    // d3 = 30 70 31 71 32 72 33 73
+    "ld1       {v31.8h}, [%7]                          \n"
+    "add       %2, %2, %0                              \n"
+  "1:                                                  \n"
+
+    // 00 40 01 41 02 42 03 43
+    // 10 50 11 51 12 52 13 53
+    // 20 60 21 61 22 62 23 63
+    // 30 70 31 71 32 72 33 73
     MEMACCESS(0)
-    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n"
+    "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32                \n"
     MEMACCESS(3)
-    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n"
+    "ld4       {v4.8b,v5.8b,v6.8b,v7.8b}, [%2], #32                \n"
     MEMACCESS(4)
-    "vld4.8       {d16, d17, d18, d19}, [%4]!  \n"
-    "subs         %2, %2, #12                  \n"
+    "ld4       {v16.8b,v17.8b,v18.8b,v19.8b}, [%3], #32              \n"
+    "subs      %4, %4, #12                             \n"
 
     // Shuffle the input data around to get align the data
     //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
-    // d0 = 00 10 01 11 02 12 03 13
-    // d1 = 40 50 41 51 42 52 43 53
-    "vtrn.u8      d0, d1                       \n"
-    "vtrn.u8      d4, d5                       \n"
-    "vtrn.u8      d16, d17                     \n"
-
-    // d2 = 20 30 21 31 22 32 23 33
-    // d3 = 60 70 61 71 62 72 63 73
-    "vtrn.u8      d2, d3                       \n"
-    "vtrn.u8      d6, d7                       \n"
-    "vtrn.u8      d18, d19                     \n"
-
-    // d0 = 00+10 01+11 02+12 03+13
-    // d2 = 40+50 41+51 42+52 43+53
-    "vpaddl.u8    q0, q0                       \n"
-    "vpaddl.u8    q2, q2                       \n"
-    "vpaddl.u8    q8, q8                       \n"
-
-    // d3 = 60+70 61+71 62+72 63+73
-    "vpaddl.u8    d3, d3                       \n"
-    "vpaddl.u8    d7, d7                       \n"
-    "vpaddl.u8    d19, d19                     \n"
+    // 00 10 01 11 02 12 03 13
+    // 40 50 41 51 42 52 43 53
+    "trn1      v20.8b, v0.8b, v1.8b                    \n"
+    "trn2      v21.8b, v0.8b, v1.8b                    \n"
+    "trn1      v22.8b, v4.8b, v5.8b                    \n"
+    "trn2      v23.8b, v4.8b, v5.8b                    \n"
+    "trn1      v24.8b, v16.8b, v17.8b                  \n"
+    "trn2      v25.8b, v16.8b, v17.8b                  \n"
+
+    // 20 30 21 31 22 32 23 33
+    // 60 70 61 71 62 72 63 73
+    "trn1      v0.8b, v2.8b, v3.8b                     \n"
+    "trn2      v1.8b, v2.8b, v3.8b                     \n"
+    "trn1      v4.8b, v6.8b, v7.8b                     \n"
+    "trn2      v5.8b, v6.8b, v7.8b                     \n"
+    "trn1      v16.8b, v18.8b, v19.8b                  \n"
+    "trn2      v17.8b, v18.8b, v19.8b                  \n"
+
+    // 00+10 01+11 02+12 03+13
+    // 40+50 41+51 42+52 43+53
+    "uaddlp    v20.4h, v20.8b                          \n"
+    "uaddlp    v21.4h, v21.8b                          \n"
+    "uaddlp    v22.4h, v22.8b                          \n"
+    "uaddlp    v23.4h, v23.8b                          \n"
+    "uaddlp    v24.4h, v24.8b                          \n"
+    "uaddlp    v25.4h, v25.8b                          \n"
+
+    // 60+70 61+71 62+72 63+73
+    "uaddlp    v1.4h, v1.8b                            \n"
+    "uaddlp    v5.4h, v5.8b                            \n"
+    "uaddlp    v17.4h, v17.8b                          \n"
 
     // combine source lines
-    "vadd.u16     q0, q2                       \n"
-    "vadd.u16     q0, q8                       \n"
-    "vadd.u16     d4, d3, d7                   \n"
-    "vadd.u16     d4, d19                      \n"
+    "add       v20.4h, v20.4h, v22.4h                  \n"
+    "add       v21.4h, v21.4h, v23.4h                  \n"
+    "add       v20.4h, v20.4h, v24.4h                  \n"
+    "add       v21.4h, v21.4h, v25.4h                  \n"
+    "add       v2.4h, v1.4h, v5.4h                     \n"
+    "add       v2.4h, v2.4h, v17.4h                    \n"
 
     // dst_ptr[3] = (s[6 + st * 0] + s[7 + st * 0]
     //             + s[6 + st * 1] + s[7 + st * 1]
     //             + s[6 + st * 2] + s[7 + st * 2]) / 6
-    "vqrdmulh.s16 q2, q2, q13                  \n"
-    "vmovn.u16    d4, q2                       \n"
+    "sqrdmulh  v2.8h, v2.8h, v29.8h                    \n"
+    "xtn       v2.8b,  v2.8h                           \n"
 
     // Shuffle 2,3 reg around so that 2 can be added to the
     //  0,1 reg and 3 can be added to the 4,5 reg. This
     //  requires expanding from u8 to u16 as the 0,1 and 4,5
     //  registers are already expanded. Then do transposes
     //  to get aligned.
-    // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
-    "vmovl.u8     q1, d2                       \n"
-    "vmovl.u8     q3, d6                       \n"
-    "vmovl.u8     q9, d18                      \n"
+    // xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
+    "ushll     v16.8h, v16.8b, #0                      \n"
+    "uaddl     v0.8h, v0.8b, v4.8b                     \n"
 
     // combine source lines
-    "vadd.u16     q1, q3                       \n"
-    "vadd.u16     q1, q9                       \n"
-
-    // d4 = xx 20 xx 30 xx 22 xx 32
-    // d5 = xx 21 xx 31 xx 23 xx 33
-    "vtrn.u32     d2, d3                       \n"
+    "add       v0.8h, v0.8h, v16.8h                    \n"
 
-    // d4 = xx 20 xx 21 xx 22 xx 23
-    // d5 = xx 30 xx 31 xx 32 xx 33
-    "vtrn.u16     d2, d3                       \n"
+    // xx 20 xx 21 xx 22 xx 23
+    // xx 30 xx 31 xx 32 xx 33
+    "trn1      v1.8h, v0.8h, v0.8h                     \n"
+    "trn2      v4.8h, v0.8h, v0.8h                     \n"
+    "xtn       v0.4h, v1.4s                            \n"
+    "xtn       v4.4h, v4.4s                            \n"
 
     // 0+1+2, 3+4+5
-    "vadd.u16     q0, q1                       \n"
+    "add       v20.8h, v20.8h, v0.8h                   \n"
+    "add       v21.8h, v21.8h, v4.8h                   \n"
 
     // Need to divide, but can't downshift as the the value
     //  isn't a power of 2. So multiply by 65536 / n
     //  and take the upper 16 bits.
-    "vqrdmulh.s16 q0, q0, q15                  \n"
+    "sqrdmulh  v0.8h, v20.8h, v31.8h                   \n"
+    "sqrdmulh  v1.8h, v21.8h, v31.8h                   \n"
 
     // Align for table lookup, vtbl requires registers to
     //  be adjacent
-    "vmov.u8      d2, d4                       \n"
-
-    "vtbl.u8      d3, {d0, d1, d2}, d28        \n"
-    "vtbl.u8      d4, {d0, d1, d2}, d29        \n"
+    "tbl       v3.16b, {v0.16b, v1.16b, v2.16b}, v30.16b \n"
 
     MEMACCESS(1)
-    "vst1.8       {d3}, [%1]!                  \n"
+    "st1       {v3.8b}, [%1], #8                       \n"
     MEMACCESS(1)
-    "vst1.32      {d4[0]}, [%1]!               \n"
-    "bgt          1b                           \n"
+    "st1       {v3.s}[2], [%1], #4                     \n"
+    "b.gt      1b                                      \n"
   : "+r"(src_ptr),          // %0
     "+r"(dst_ptr),          // %1
-    "+r"(dst_width),        // %2
-    "+r"(src_stride),       // %3
-    "+r"(src_ptr1)          // %4
+    "+r"(tmp_src_stride),   // %2
+    "+r"(src_ptr1),         // %3
+    "+r"(dst_width)         // %4
   : "r"(&kMult38_Div6),     // %5
     "r"(&kShuf38_2),        // %6
     "r"(&kMult38_Div9)      // %7
-  : "q0", "q1", "q2", "q3", "q8", "q9", "q13", "q14", "q15", "memory", "cc"
+  : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17",
+    "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v29",
+    "v30", "v31", "memory", "cc"
   );
 }
-#endif //HAS_SCALEROWDOWN38_NEON
 
-#ifdef HAS_SCALEROWDOWN38_NEON
 // 32x2 -> 12x1
 void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr,
                                ptrdiff_t src_stride,
                                uint8* dst_ptr, int dst_width) {
+  // TODO(fbarchard): use src_stride directly for clang 3.5+.
+  ptrdiff_t tmp_src_stride = src_stride;
   asm volatile (
     MEMACCESS(4)
-    "vld1.16    {q13}, [%4]                    \n"
+    "ld1       {v30.8h}, [%4]                          \n"
     MEMACCESS(5)
-    "vld1.8     {q14}, [%5]                    \n"
-    "add        %3, %0                         \n"
-    ".p2align   2                              \n"
-  "1:                                          \n"
-
-    // d0 = 00 40 01 41 02 42 03 43
-    // d1 = 10 50 11 51 12 52 13 53
-    // d2 = 20 60 21 61 22 62 23 63
-    // d3 = 30 70 31 71 32 72 33 73
+    "ld1       {v31.16b}, [%5]                         \n"
+    "add       %2, %2, %0                              \n"
+  "1:                                                  \n"
+
+    // 00 40 01 41 02 42 03 43
+    // 10 50 11 51 12 52 13 53
+    // 20 60 21 61 22 62 23 63
+    // 30 70 31 71 32 72 33 73
     MEMACCESS(0)
-    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n"
+    "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32                \n"
     MEMACCESS(3)
-    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n"
-    "subs         %2, %2, #12                  \n"
+    "ld4       {v4.8b,v5.8b,v6.8b,v7.8b}, [%2], #32                \n"
+    "subs      %3, %3, #12                             \n"
 
     // Shuffle the input data around to get align the data
     //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
-    // d0 = 00 10 01 11 02 12 03 13
-    // d1 = 40 50 41 51 42 52 43 53
-    "vtrn.u8      d0, d1                       \n"
-    "vtrn.u8      d4, d5                       \n"
-
-    // d2 = 20 30 21 31 22 32 23 33
-    // d3 = 60 70 61 71 62 72 63 73
-    "vtrn.u8      d2, d3                       \n"
-    "vtrn.u8      d6, d7                       \n"
-
-    // d0 = 00+10 01+11 02+12 03+13
-    // d2 = 40+50 41+51 42+52 43+53
-    "vpaddl.u8    q0, q0                       \n"
-    "vpaddl.u8    q2, q2                       \n"
-
-    // d3 = 60+70 61+71 62+72 63+73
-    "vpaddl.u8    d3, d3                       \n"
-    "vpaddl.u8    d7, d7                       \n"
+    // 00 10 01 11 02 12 03 13
+    // 40 50 41 51 42 52 43 53
+    "trn1      v16.8b, v0.8b, v1.8b                    \n"
+    "trn2      v17.8b, v0.8b, v1.8b                    \n"
+    "trn1      v18.8b, v4.8b, v5.8b                    \n"
+    "trn2      v19.8b, v4.8b, v5.8b                    \n"
+
+    // 20 30 21 31 22 32 23 33
+    // 60 70 61 71 62 72 63 73
+    "trn1      v0.8b, v2.8b, v3.8b                     \n"
+    "trn2      v1.8b, v2.8b, v3.8b                     \n"
+    "trn1      v4.8b, v6.8b, v7.8b                     \n"
+    "trn2      v5.8b, v6.8b, v7.8b                     \n"
+
+    // 00+10 01+11 02+12 03+13
+    // 40+50 41+51 42+52 43+53
+    "uaddlp    v16.4h, v16.8b                          \n"
+    "uaddlp    v17.4h, v17.8b                          \n"
+    "uaddlp    v18.4h, v18.8b                          \n"
+    "uaddlp    v19.4h, v19.8b                          \n"
+
+    // 60+70 61+71 62+72 63+73
+    "uaddlp    v1.4h, v1.8b                            \n"
+    "uaddlp    v5.4h, v5.8b                            \n"
 
     // combine source lines
-    "vadd.u16     q0, q2                       \n"
-    "vadd.u16     d4, d3, d7                   \n"
+    "add       v16.4h, v16.4h, v18.4h                  \n"
+    "add       v17.4h, v17.4h, v19.4h                  \n"
+    "add       v2.4h, v1.4h, v5.4h                     \n"
 
     // dst_ptr[3] = (s[6] + s[7] + s[6+st] + s[7+st]) / 4
-    "vqrshrn.u16  d4, q2, #2                   \n"
+    "uqrshrn   v2.8b, v2.8h, #2                        \n"
 
     // Shuffle 2,3 reg around so that 2 can be added to the
     //  0,1 reg and 3 can be added to the 4,5 reg. This
     //  requires expanding from u8 to u16 as the 0,1 and 4,5
     //  registers are already expanded. Then do transposes
     //  to get aligned.
-    // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
-    "vmovl.u8     q1, d2                       \n"
-    "vmovl.u8     q3, d6                       \n"
+    // xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
 
     // combine source lines
-    "vadd.u16     q1, q3                       \n"
-
-    // d4 = xx 20 xx 30 xx 22 xx 32
-    // d5 = xx 21 xx 31 xx 23 xx 33
-    "vtrn.u32     d2, d3                       \n"
+    "uaddl     v0.8h, v0.8b, v4.8b                     \n"
 
-    // d4 = xx 20 xx 21 xx 22 xx 23
-    // d5 = xx 30 xx 31 xx 32 xx 33
-    "vtrn.u16     d2, d3                       \n"
+    // xx 20 xx 21 xx 22 xx 23
+    // xx 30 xx 31 xx 32 xx 33
+    "trn1      v1.8h, v0.8h, v0.8h                     \n"
+    "trn2      v4.8h, v0.8h, v0.8h                     \n"
+    "xtn       v0.4h, v1.4s                            \n"
+    "xtn       v4.4h, v4.4s                            \n"
 
     // 0+1+2, 3+4+5
-    "vadd.u16     q0, q1                       \n"
+    "add       v16.8h, v16.8h, v0.8h                   \n"
+    "add       v17.8h, v17.8h, v4.8h                   \n"
 
     // Need to divide, but can't downshift as the the value
     //  isn't a power of 2. So multiply by 65536 / n
     //  and take the upper 16 bits.
-    "vqrdmulh.s16 q0, q0, q13                  \n"
+    "sqrdmulh  v0.8h, v16.8h, v30.8h                   \n"
+    "sqrdmulh  v1.8h, v17.8h, v30.8h                   \n"
 
     // Align for table lookup, vtbl requires registers to
     //  be adjacent
-    "vmov.u8      d2, d4                       \n"
 
-    "vtbl.u8      d3, {d0, d1, d2}, d28        \n"
-    "vtbl.u8      d4, {d0, d1, d2}, d29        \n"
+    "tbl       v3.16b, {v0.16b, v1.16b, v2.16b}, v31.16b \n"
 
     MEMACCESS(1)
-    "vst1.8       {d3}, [%1]!                  \n"
+    "st1       {v3.8b}, [%1], #8                       \n"
     MEMACCESS(1)
-    "vst1.32      {d4[0]}, [%1]!               \n"
-    "bgt          1b                           \n"
-  : "+r"(src_ptr),       // %0
-    "+r"(dst_ptr),       // %1
-    "+r"(dst_width),     // %2
-    "+r"(src_stride)     // %3
-  : "r"(&kMult38_Div6),  // %4
-    "r"(&kShuf38_2)      // %5
-  : "q0", "q1", "q2", "q3", "q13", "q14", "memory", "cc"
+    "st1       {v3.s}[2], [%1], #4                     \n"
+    "b.gt      1b                                      \n"
+  : "+r"(src_ptr),         // %0
+    "+r"(dst_ptr),         // %1
+    "+r"(tmp_src_stride),  // %2
+    "+r"(dst_width)        // %3
+  : "r"(&kMult38_Div6),    // %4
+    "r"(&kShuf38_2)        // %5
+  : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17",
+    "v18", "v19", "v30", "v31", "memory", "cc"
   );
 }
-#endif //HAS_SCALEROWDOWN38_NEON
 
-#if 0
 // 16x2 -> 16x1
 void ScaleFilterRows_NEON(uint8* dst_ptr,
                           const uint8* src_ptr, ptrdiff_t src_stride,
                           int dst_width, int source_y_fraction) {
+    int y_fraction = 256 - source_y_fraction;
   asm volatile (
     "cmp          %4, #0                       \n"
-    "beq          100f                         \n"
-    "add          %2, %1                       \n"
+    "b.eq         100f                         \n"
+    "add          %2, %2, %1                   \n"
     "cmp          %4, #64                      \n"
-    "beq          75f                          \n"
+    "b.eq         75f                          \n"
     "cmp          %4, #128                     \n"
-    "beq          50f                          \n"
+    "b.eq         50f                          \n"
     "cmp          %4, #192                     \n"
-    "beq          25f                          \n"
+    "b.eq         25f                          \n"
 
-    "vdup.8       d5, %4                       \n"
-    "rsb          %4, #256                     \n"
-    "vdup.8       d4, %4                       \n"
+    "dup          v5.8b, %w4                   \n"
+    "dup          v4.8b, %w5                   \n"
     // General purpose row blend.
   "1:                                          \n"
     MEMACCESS(1)
-    "vld1.8       {q0}, [%1]!                  \n"
+    "ld1          {v0.16b}, [%1], #16          \n"
     MEMACCESS(2)
-    "vld1.8       {q1}, [%2]!                  \n"
+    "ld1          {v1.16b}, [%2], #16          \n"
     "subs         %3, %3, #16                  \n"
-    "vmull.u8     q13, d0, d4                  \n"
-    "vmull.u8     q14, d1, d4                  \n"
-    "vmlal.u8     q13, d2, d5                  \n"
-    "vmlal.u8     q14, d3, d5                  \n"
-    "vrshrn.u16   d0, q13, #8                  \n"
-    "vrshrn.u16   d1, q14, #8                  \n"
+    "umull        v6.8h, v0.8b, v4.8b          \n"
+    "umull2       v7.8h, v0.16b, v4.16b        \n"
+    "umlal        v6.8h, v1.8b, v5.8b          \n"
+    "umlal2       v7.8h, v1.16b, v5.16b        \n"
+    "rshrn        v0.8b, v6.8h, #8             \n"
+    "rshrn2       v0.16b, v7.8h, #8            \n"
     MEMACCESS(0)
-    "vst1.8       {q0}, [%0]!                  \n"
-    "bgt          1b                           \n"
+    "st1          {v0.16b}, [%0], #16          \n"
+    "b.gt         1b                           \n"
     "b            99f                          \n"
 
     // Blend 25 / 75.
   "25:                                         \n"
     MEMACCESS(1)
-    "vld1.8       {q0}, [%1]!                  \n"
+    "ld1          {v0.16b}, [%1], #16          \n"
     MEMACCESS(2)
-    "vld1.8       {q1}, [%2]!                  \n"
+    "ld1          {v1.16b}, [%2], #16          \n"
     "subs         %3, %3, #16                  \n"
-    "vrhadd.u8    q0, q1                       \n"
-    "vrhadd.u8    q0, q1                       \n"
+    "urhadd       v0.16b, v0.16b, v1.16b       \n"
+    "urhadd       v0.16b, v0.16b, v1.16b       \n"
     MEMACCESS(0)
-    "vst1.8       {q0}, [%0]!                  \n"
-    "bgt          25b                          \n"
+    "st1          {v0.16b}, [%0], #16          \n"
+    "b.gt         25b                          \n"
     "b            99f                          \n"
 
     // Blend 50 / 50.
   "50:                                         \n"
     MEMACCESS(1)
-    "vld1.8       {q0}, [%1]!                  \n"
+    "ld1          {v0.16b}, [%1], #16          \n"
     MEMACCESS(2)
-    "vld1.8       {q1}, [%2]!                  \n"
+    "ld1          {v1.16b}, [%2], #16          \n"
     "subs         %3, %3, #16                  \n"
-    "vrhadd.u8    q0, q1                       \n"
+    "urhadd       v0.16b, v0.16b, v1.16b       \n"
     MEMACCESS(0)
-    "vst1.8       {q0}, [%0]!                  \n"
-    "bgt          50b                          \n"
+    "st1          {v0.16b}, [%0], #16          \n"
+    "b.gt         50b                          \n"
     "b            99f                          \n"
 
     // Blend 75 / 25.
   "75:                                         \n"
     MEMACCESS(1)
-    "vld1.8       {q1}, [%1]!                  \n"
+    "ld1          {v1.16b}, [%1], #16          \n"
     MEMACCESS(2)
-    "vld1.8       {q0}, [%2]!                  \n"
+    "ld1          {v0.16b}, [%2], #16          \n"
     "subs         %3, %3, #16                  \n"
-    "vrhadd.u8    q0, q1                       \n"
-    "vrhadd.u8    q0, q1                       \n"
+    "urhadd       v0.16b, v0.16b, v1.16b       \n"
+    "urhadd       v0.16b, v0.16b, v1.16b       \n"
     MEMACCESS(0)
-    "vst1.8       {q0}, [%0]!                  \n"
-    "bgt          75b                          \n"
+    "st1          {v0.16b}, [%0], #16          \n"
+    "b.gt         75b                          \n"
     "b            99f                          \n"
 
     // Blend 100 / 0 - Copy row unchanged.
   "100:                                        \n"
     MEMACCESS(1)
-    "vld1.8       {q0}, [%1]!                  \n"
+    "ld1          {v0.16b}, [%1], #16          \n"
     "subs         %3, %3, #16                  \n"
     MEMACCESS(0)
-    "vst1.8       {q0}, [%0]!                  \n"
-    "bgt          100b                         \n"
+    "st1          {v0.16b}, [%0], #16          \n"
+    "b.gt         100b                         \n"
 
   "99:                                         \n"
     MEMACCESS(0)
-    "vst1.8       {d1[7]}, [%0]                \n"
+    "st1          {v0.b}[15], [%0]             \n"
   : "+r"(dst_ptr),          // %0
     "+r"(src_ptr),          // %1
     "+r"(src_stride),       // %2
     "+r"(dst_width),        // %3
-    "+r"(source_y_fraction) // %4
+    "+r"(source_y_fraction),// %4
+    "+r"(y_fraction)        // %5
   :
-  : "q0", "q1", "d4", "d5", "q13", "q14", "memory", "cc"
+  : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "memory", "cc"
   );
 }
-#endif //0
 
-#ifdef HAS_SCALEARGBROWDOWN2_NEON
 void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
                             uint8* dst, int dst_width) {
   asm volatile (
-    ".p2align   2                              \n"
   "1:                                          \n"
     // load even pixels into q0, odd into q1
-    MEMACCESS(0)
-    "vld2.32    {q0, q1}, [%0]!                \n"
-    MEMACCESS(0)
-    "vld2.32    {q2, q3}, [%0]!                \n"
+    MEMACCESS (0)
+    "ld2        {v0.4s, v1.4s}, [%0], #32      \n"
+    MEMACCESS (0)
+    "ld2        {v2.4s, v3.4s}, [%0], #32      \n"
     "subs       %2, %2, #8                     \n"  // 8 processed per loop
-    MEMACCESS(1)
-    "vst1.8     {q1}, [%1]!                    \n"  // store odd pixels
-    MEMACCESS(1)
-    "vst1.8     {q3}, [%1]!                    \n"
-    "bgt        1b                             \n"
-  : "+r"(src_ptr),          // %0
-    "+r"(dst),              // %1
-    "+r"(dst_width)         // %2
+    MEMACCESS (1)
+    "st1        {v1.16b}, [%1], #16            \n"  // store odd pixels
+    MEMACCESS (1)
+    "st1        {v3.16b}, [%1], #16            \n"
+    "b.gt       1b                             \n"
+  : "+r" (src_ptr),          // %0
+    "+r" (dst),              // %1
+    "+r" (dst_width)         // %2
   :
-  : "memory", "cc", "q0", "q1", "q2", "q3"  // Clobber List
+  : "memory", "cc", "v0", "v1", "v2", "v3"  // Clobber List
   );
 }
-#endif //HAS_SCALEARGBROWDOWN2_NEON
 
-#ifdef HAS_SCALEARGBROWDOWN2_NEON
 void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
                                uint8* dst, int dst_width) {
   asm volatile (
     // change the stride to row 2 pointer
     "add        %1, %1, %0                     \n"
-    ".p2align   2                              \n"
   "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
-    MEMACCESS(0)
-    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    MEMACCESS (0)
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64   \n"  // load 8 ARGB pixels.
     "subs       %3, %3, #8                     \n"  // 8 processed per loop.
-    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
-    "vpaddl.u8  q3, q3                         \n"  // A 16 bytes -> 8 shorts.
-    MEMACCESS(1)
-    "vld4.8     {d16, d18, d20, d22}, [%1]!    \n"  // load 8 more ARGB pixels.
-    MEMACCESS(1)
-    "vld4.8     {d17, d19, d21, d23}, [%1]!    \n"  // load last 8 ARGB pixels.
-    "vpadal.u8  q0, q8                         \n"  // B 16 bytes -> 8 shorts.
-    "vpadal.u8  q1, q9                         \n"  // G 16 bytes -> 8 shorts.
-    "vpadal.u8  q2, q10                        \n"  // R 16 bytes -> 8 shorts.
-    "vpadal.u8  q3, q11                        \n"  // A 16 bytes -> 8 shorts.
-    "vrshrn.u16 d0, q0, #2                     \n"  // downshift, round and pack
-    "vrshrn.u16 d1, q1, #2                     \n"
-    "vrshrn.u16 d2, q2, #2                     \n"
-    "vrshrn.u16 d3, q3, #2                     \n"
-    MEMACCESS(2)
-    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"
-    "bgt        1b                             \n"
-  : "+r"(src_ptr),          // %0
-    "+r"(src_stride),       // %1
-    "+r"(dst),              // %2
-    "+r"(dst_width)         // %3
+    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
+    "uaddlp     v3.8h, v3.16b                  \n"  // A 16 bytes -> 8 shorts.
+    MEMACCESS (1)
+    "ld4        {v16.16b,v17.16b,v18.16b,v19.16b}, [%1], #64 \n"  // load 8 more ARGB pixels.
+    "uadalp     v0.8h, v16.16b                 \n"  // B 16 bytes -> 8 shorts.
+    "uadalp     v1.8h, v17.16b                 \n"  // G 16 bytes -> 8 shorts.
+    "uadalp     v2.8h, v18.16b                 \n"  // R 16 bytes -> 8 shorts.
+    "uadalp     v3.8h, v19.16b                 \n"  // A 16 bytes -> 8 shorts.
+    "rshrn      v0.8b, v0.8h, #2               \n"  // downshift, round and pack
+    "rshrn      v1.8b, v1.8h, #2               \n"
+    "rshrn      v2.8b, v2.8h, #2               \n"
+    "rshrn      v3.8b, v3.8h, #2               \n"
+    MEMACCESS (2)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32     \n"
+    "b.gt       1b                             \n"
+  : "+r" (src_ptr),          // %0
+    "+r" (src_stride),       // %1
+    "+r" (dst),              // %2
+    "+r" (dst_width)         // %3
   :
-  : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11"
+  : "memory", "cc", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19"
   );
 }
-#endif //HAS_SCALEARGBROWDOWN2_NEON
 
-#ifdef HAS_SCALEARGBROWDOWNEVEN_NEON
 // Reads 4 pixels at a time.
 // Alignment requirement: src_argb 4 byte aligned.
 void ScaleARGBRowDownEven_NEON(const uint8* src_argb,  ptrdiff_t src_stride,
                                int src_stepx, uint8* dst_argb, int dst_width) {
   asm volatile (
-    "mov        r12, %3, lsl #2                \n"
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "vld1.32    {d0[0]}, [%0], r12             \n"
+    "ld1        {v0.s}[0], [%0], %3            \n"
     MEMACCESS(0)
-    "vld1.32    {d0[1]}, [%0], r12             \n"
+    "ld1        {v0.s}[1], [%0], %3            \n"
     MEMACCESS(0)
-    "vld1.32    {d1[0]}, [%0], r12             \n"
+    "ld1        {v0.s}[2], [%0], %3            \n"
     MEMACCESS(0)
-    "vld1.32    {d1[1]}, [%0], r12             \n"
+    "ld1        {v0.s}[3], [%0], %3            \n"
     "subs       %2, %2, #4                     \n"  // 4 pixels per loop.
     MEMACCESS(1)
-    "vst1.8     {q0}, [%1]!                    \n"
-    "bgt        1b                             \n"
+    "st1        {v0.16b}, [%1], #16            \n"
+    "b.gt       1b                             \n"
   : "+r"(src_argb),    // %0
     "+r"(dst_argb),    // %1
     "+r"(dst_width)    // %2
-  : "r"(src_stepx)     // %3
-  : "memory", "cc", "r12", "q0"
+  : "r"(static_cast<ptrdiff_t>(src_stepx * 4)) // %3
+  : "memory", "cc", "v0"
   );
 }
-#endif //HAS_SCALEARGBROWDOWNEVEN_NEON
 
-#ifdef HAS_SCALEARGBROWDOWNEVEN_NEON
 // Reads 4 pixels at a time.
 // Alignment requirement: src_argb 4 byte aligned.
+// TODO, might be worth another optimization pass in future.
+// It could be upgraded to 8 pixels at a time to start with.
 void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, ptrdiff_t src_stride,
                                   int src_stepx,
                                   uint8* dst_argb, int dst_width) {
   asm volatile (
-    "mov        r12, %4, lsl #2                \n"
     "add        %1, %1, %0                     \n"
-    ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "vld1.8     {d0}, [%0], r12                \n"  // Read 4 2x2 blocks -> 2x1
+    "ld1     {v0.8b}, [%0], %4                 \n"  // Read 4 2x2 blocks -> 2x1
     MEMACCESS(1)
-    "vld1.8     {d1}, [%1], r12                \n"
+    "ld1     {v1.8b}, [%1], %4                 \n"
     MEMACCESS(0)
-    "vld1.8     {d2}, [%0], r12                \n"
+    "ld1     {v2.8b}, [%0], %4                 \n"
     MEMACCESS(1)
-    "vld1.8     {d3}, [%1], r12                \n"
+    "ld1     {v3.8b}, [%1], %4                 \n"
     MEMACCESS(0)
-    "vld1.8     {d4}, [%0], r12                \n"
+    "ld1     {v4.8b}, [%0], %4                 \n"
     MEMACCESS(1)
-    "vld1.8     {d5}, [%1], r12                \n"
+    "ld1     {v5.8b}, [%1], %4                 \n"
     MEMACCESS(0)
-    "vld1.8     {d6}, [%0], r12                \n"
+    "ld1     {v6.8b}, [%0], %4                 \n"
     MEMACCESS(1)
-    "vld1.8     {d7}, [%1], r12                \n"
-    "vaddl.u8   q0, d0, d1                     \n"
-    "vaddl.u8   q1, d2, d3                     \n"
-    "vaddl.u8   q2, d4, d5                     \n"
-    "vaddl.u8   q3, d6, d7                     \n"
-    "vswp.8     d1, d2                         \n"  // ab_cd -> ac_bd
-    "vswp.8     d5, d6                         \n"  // ef_gh -> eg_fh
-    "vadd.u16   q0, q0, q1                     \n"  // (a+b)_(c+d)
-    "vadd.u16   q2, q2, q3                     \n"  // (e+f)_(g+h)
-    "vrshrn.u16 d0, q0, #2                     \n"  // first 2 pixels.
-    "vrshrn.u16 d1, q2, #2                     \n"  // next 2 pixels.
+    "ld1     {v7.8b}, [%1], %4                 \n"
+    "uaddl   v0.8h, v0.8b, v1.8b               \n"
+    "uaddl   v2.8h, v2.8b, v3.8b               \n"
+    "uaddl   v4.8h, v4.8b, v5.8b               \n"
+    "uaddl   v6.8h, v6.8b, v7.8b               \n"
+    "mov     v16.d[1], v0.d[1]                 \n"  // ab_cd -> ac_bd
+    "mov     v0.d[1], v2.d[0]                  \n"
+    "mov     v2.d[0], v16.d[1]                 \n"
+    "mov     v16.d[1], v4.d[1]                 \n"  // ef_gh -> eg_fh
+    "mov     v4.d[1], v6.d[0]                  \n"
+    "mov     v6.d[0], v16.d[1]                 \n"
+    "add     v0.8h, v0.8h, v2.8h               \n"  // (a+b)_(c+d)
+    "add     v4.8h, v4.8h, v6.8h               \n"  // (e+f)_(g+h)
+    "rshrn   v0.8b, v0.8h, #2                  \n"  // first 2 pixels.
+    "rshrn2  v0.16b, v4.8h, #2                 \n"  // next 2 pixels.
     "subs       %3, %3, #4                     \n"  // 4 pixels per loop.
     MEMACCESS(2)
-    "vst1.8     {q0}, [%2]!                    \n"
-    "bgt        1b                             \n"
+    "st1     {v0.16b}, [%2], #16               \n"
+    "b.gt       1b                             \n"
   : "+r"(src_argb),    // %0
     "+r"(src_stride),  // %1
     "+r"(dst_argb),    // %2
     "+r"(dst_width)    // %3
-  : "r"(src_stepx)     // %4
-  : "memory", "cc", "r12", "q0", "q1", "q2", "q3"
+  : "r"(src_stepx * 4) // %4
+  : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"
   );
 }
-#endif  // HAS_SCALEARGBROWDOWNEVEN_NEON
-#endif  // __aarch64__
+#endif  // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
 
 #ifdef __cplusplus
 }  // extern "C"
diff --git a/third_party/libyuv/source/scale_posix.cc b/third_party/libyuv/source/scale_posix.cc
index 352e66782..bb6e57efe 100644
--- a/third_party/libyuv/source/scale_posix.cc
+++ b/third_party/libyuv/source/scale_posix.cc
@@ -101,110 +101,6 @@ void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
   asm volatile (
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "psrlw     $0x8,%%xmm0                     \n"
-    "psrlw     $0x8,%%xmm1                     \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_ptr),    // %0
-    "+r"(dst_ptr),    // %1
-    "+r"(dst_width)   // %2
-  :
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1"
-#endif
-  );
-}
-
-void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
-                              uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "psrlw     $0x8,%%xmm5                     \n"
-
-    LABELALIGN
-  "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10, 0) ",%%xmm1  \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "movdqa    %%xmm0,%%xmm2                   \n"
-    "psrlw     $0x8,%%xmm0                     \n"
-    "movdqa    %%xmm1,%%xmm3                   \n"
-    "psrlw     $0x8,%%xmm1                     \n"
-    "pand      %%xmm5,%%xmm2                   \n"
-    "pand      %%xmm5,%%xmm3                   \n"
-    "pavgw     %%xmm2,%%xmm0                   \n"
-    "pavgw     %%xmm3,%%xmm1                   \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_ptr),    // %0
-    "+r"(dst_ptr),    // %1
-    "+r"(dst_width)   // %2
-  :
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm5"
-#endif
-  );
-}
-
-void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
-                           uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "psrlw     $0x8,%%xmm5                     \n"
-
-    LABELALIGN
-  "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    MEMOPREG(movdqa,0x00,0,3,1,xmm2)           //  movdqa  (%0,%3,1),%%xmm2
-    BUNDLEALIGN
-    MEMOPREG(movdqa,0x10,0,3,1,xmm3)           //  movdqa  0x10(%0,%3,1),%%xmm3
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "pavgb     %%xmm2,%%xmm0                   \n"
-    "pavgb     %%xmm3,%%xmm1                   \n"
-    "movdqa    %%xmm0,%%xmm2                   \n"
-    "psrlw     $0x8,%%xmm0                     \n"
-    "movdqa    %%xmm1,%%xmm3                   \n"
-    "psrlw     $0x8,%%xmm1                     \n"
-    "pand      %%xmm5,%%xmm2                   \n"
-    "pand      %%xmm5,%%xmm3                   \n"
-    "pavgw     %%xmm2,%%xmm0                   \n"
-    "pavgw     %%xmm3,%%xmm1                   \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_ptr),    // %0
-    "+r"(dst_ptr),    // %1
-    "+r"(dst_width)   // %2
-  : "r"((intptr_t)(src_stride))   // %3
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-#endif
-  );
-}
-
-void ScaleRowDown2_Unaligned_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
-                                  uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    LABELALIGN
-  "1:                                          \n"
     "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
     "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
     "lea       " MEMLEA(0x20,0) ",%0           \n"
@@ -218,17 +114,12 @@ void ScaleRowDown2_Unaligned_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
   : "+r"(src_ptr),    // %0
     "+r"(dst_ptr),    // %1
     "+r"(dst_width)   // %2
-  :
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1"
-#endif
+  :: "memory", "cc", "xmm0", "xmm1"
   );
 }
 
-void ScaleRowDown2Linear_Unaligned_SSE2(const uint8* src_ptr,
-                                        ptrdiff_t src_stride,
-                                        uint8* dst_ptr, int dst_width) {
+void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width) {
   asm volatile (
     "pcmpeqb   %%xmm5,%%xmm5                   \n"
     "psrlw     $0x8,%%xmm5                     \n"
@@ -236,7 +127,7 @@ void ScaleRowDown2Linear_Unaligned_SSE2(const uint8* src_ptr,
     LABELALIGN
   "1:                                          \n"
     "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x10, 0) ",%%xmm1  \n"
     "lea       " MEMLEA(0x20,0) ",%0           \n"
     "movdqa    %%xmm0,%%xmm2                   \n"
     "psrlw     $0x8,%%xmm0                     \n"
@@ -254,17 +145,12 @@ void ScaleRowDown2Linear_Unaligned_SSE2(const uint8* src_ptr,
   : "+r"(src_ptr),    // %0
     "+r"(dst_ptr),    // %1
     "+r"(dst_width)   // %2
-  :
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm5"
-#endif
+  :: "memory", "cc", "xmm0", "xmm1", "xmm5"
   );
 }
 
-void ScaleRowDown2Box_Unaligned_SSE2(const uint8* src_ptr,
-                                     ptrdiff_t src_stride,
-                                     uint8* dst_ptr, int dst_width) {
+void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
   asm volatile (
     "pcmpeqb   %%xmm5,%%xmm5                   \n"
     "psrlw     $0x8,%%xmm5                     \n"
@@ -274,7 +160,6 @@ void ScaleRowDown2Box_Unaligned_SSE2(const uint8* src_ptr,
     "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
     "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
     MEMOPREG(movdqu,0x00,0,3,1,xmm2)           //  movdqu  (%0,%3,1),%%xmm2
-    BUNDLEALIGN
     MEMOPREG(movdqu,0x10,0,3,1,xmm3)           //  movdqu  0x10(%0,%3,1),%%xmm3
     "lea       " MEMLEA(0x20,0) ",%0           \n"
     "pavgb     %%xmm2,%%xmm0                   \n"
@@ -296,13 +181,8 @@ void ScaleRowDown2Box_Unaligned_SSE2(const uint8* src_ptr,
     "+r"(dst_ptr),    // %1
     "+r"(dst_width)   // %2
   : "r"((intptr_t)(src_stride))   // %3
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
   );
 }
 
@@ -315,8 +195,8 @@ void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
 
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
     "lea       " MEMLEA(0x20,0) ",%0           \n"
     "pand      %%xmm5,%%xmm0                   \n"
     "pand      %%xmm5,%%xmm1                   \n"
@@ -330,11 +210,7 @@ void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
   : "+r"(src_ptr),    // %0
     "+r"(dst_ptr),    // %1
     "+r"(dst_width)   // %2
-  :
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm5"
-#endif
+  :: "memory", "cc", "xmm0", "xmm1", "xmm5"
   );
 }
 
@@ -348,18 +224,16 @@ void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
 
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    MEMOPREG(movdqa,0x00,0,4,1,xmm2)           //  movdqa  (%0,%4,1),%%xmm2
-    BUNDLEALIGN
-    MEMOPREG(movdqa,0x10,0,4,1,xmm3)           //  movdqa  0x10(%0,%4,1),%%xmm3
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x00,0,4,1,xmm2)           //  movdqu  (%0,%4,1),%%xmm2
+    MEMOPREG(movdqu,0x10,0,4,1,xmm3)           //  movdqu  0x10(%0,%4,1),%%xmm3
     "pavgb     %%xmm2,%%xmm0                   \n"
     "pavgb     %%xmm3,%%xmm1                   \n"
-    MEMOPREG(movdqa,0x00,0,4,2,xmm2)           //  movdqa  (%0,%4,2),%%xmm2
-    BUNDLEALIGN
-    MEMOPREG(movdqa,0x10,0,4,2,xmm3)           //  movdqa  0x10(%0,%4,2),%%xmm3
-    MEMOPREG(movdqa,0x00,0,3,1,xmm4)           //  movdqa  (%0,%3,1),%%xmm4
-    MEMOPREG(movdqa,0x10,0,3,1,xmm5)           //  movdqa  0x10(%0,%3,1),%%xmm5
+    MEMOPREG(movdqu,0x00,0,4,2,xmm2)           //  movdqu  (%0,%4,2),%%xmm2
+    MEMOPREG(movdqu,0x10,0,4,2,xmm3)           //  movdqu  0x10(%0,%4,2),%%xmm3
+    MEMOPREG(movdqu,0x00,0,3,1,xmm4)           //  movdqu  (%0,%3,1),%%xmm4
+    MEMOPREG(movdqu,0x10,0,3,1,xmm5)           //  movdqu  0x10(%0,%3,1),%%xmm5
     "lea       " MEMLEA(0x20,0) ",%0           \n"
     "pavgb     %%xmm4,%%xmm2                   \n"
     "pavgb     %%xmm2,%%xmm0                   \n"
@@ -388,13 +262,8 @@ void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
     "+r"(dst_width),   // %2
     "+r"(stridex3)     // %3
   : "r"((intptr_t)(src_stride))    // %4
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm7"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm7"
   );
 }
 
@@ -412,8 +281,8 @@ void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
   asm volatile (
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm2   \n"
     "lea       " MEMLEA(0x20,0) ",%0           \n"
     "movdqa    %%xmm2,%%xmm1                   \n"
     "palignr   $0x8,%%xmm0,%%xmm1              \n"
@@ -429,11 +298,7 @@ void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
   : "+r"(src_ptr),   // %0
     "+r"(dst_ptr),   // %1
     "+r"(dst_width)  // %2
-  :
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-#endif
+  :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
 }
 
@@ -461,8 +326,8 @@ void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
   asm volatile (
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm6         \n"
-    MEMOPREG(movdqa,0x00,0,3,1,xmm7)           //  movdqa  (%0,%3),%%xmm7
+    "movdqu    " MEMACCESS(0) ",%%xmm6         \n"
+    MEMOPREG(movdqu,0x00,0,3,1,xmm7)           //  movdqu  (%0,%3),%%xmm7
     "pavgb     %%xmm7,%%xmm6                   \n"
     "pshufb    %%xmm2,%%xmm6                   \n"
     "pmaddubsw %%xmm5,%%xmm6                   \n"
@@ -479,9 +344,8 @@ void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
     "psrlw     $0x2,%%xmm6                     \n"
     "packuswb  %%xmm6,%%xmm6                   \n"
     "movq      %%xmm6," MEMACCESS2(0x8,1) "    \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
-    BUNDLEALIGN
-    MEMOPREG(movdqa,0x10,0,3,1,xmm7)           //  movdqa  0x10(%0,%3),%%xmm7
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
+    MEMOPREG(movdqu,0x10,0,3,1,xmm7)           //  movdqu  0x10(%0,%3),%%xmm7
     "lea       " MEMLEA(0x20,0) ",%0           \n"
     "pavgb     %%xmm7,%%xmm6                   \n"
     "pshufb    %%xmm4,%%xmm6                   \n"
@@ -498,13 +362,8 @@ void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
     "+r"(dst_width)  // %2
   : "r"((intptr_t)(src_stride)),  // %3
     "m"(kMadd21)     // %4
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
   );
 }
 
@@ -533,8 +392,8 @@ void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
   asm volatile (
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm6         \n"
-    MEMOPREG(movdqa,0x00,0,3,1,xmm7)           //  movdqa  (%0,%3,1),%%xmm7
+    "movdqu    " MEMACCESS(0) ",%%xmm6         \n"
+    MEMOPREG(movdqu,0x00,0,3,1,xmm7)           //  movdqu  (%0,%3,1),%%xmm7
     "pavgb     %%xmm6,%%xmm7                   \n"
     "pavgb     %%xmm7,%%xmm6                   \n"
     "pshufb    %%xmm2,%%xmm6                   \n"
@@ -553,8 +412,8 @@ void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
     "psrlw     $0x2,%%xmm6                     \n"
     "packuswb  %%xmm6,%%xmm6                   \n"
     "movq      %%xmm6," MEMACCESS2(0x8,1) "    \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
-    MEMOPREG(movdqa,0x10,0,3,1,xmm7)           //  movdqa  0x10(%0,%3,1),%%xmm7
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
+    MEMOPREG(movdqu,0x10,0,3,1,xmm7)           //  movdqu  0x10(%0,%3,1),%%xmm7
     "lea       " MEMLEA(0x20,0) ",%0           \n"
     "pavgb     %%xmm6,%%xmm7                   \n"
     "pavgb     %%xmm7,%%xmm6                   \n"
@@ -572,13 +431,8 @@ void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
       "+r"(dst_width)  // %2
     : "r"((intptr_t)(src_stride)),  // %3
       "m"(kMadd21)     // %4
-    : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-#endif
+    : "memory", "cc", NACL_R14
+      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
   );
 }
 
@@ -590,8 +444,8 @@ void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
 
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
     "lea       " MEMLEA(0x20,0) ",%0           \n"
     "pshufb    %%xmm4,%%xmm0                   \n"
     "pshufb    %%xmm5,%%xmm1                   \n"
@@ -607,10 +461,7 @@ void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
     "+r"(dst_width)  // %2
   : "m"(kShuf38a),   // %3
     "m"(kShuf38b)    // %4
-  : "memory", "cc"
-#if defined(__SSE2__)
-      , "xmm0", "xmm1", "xmm4", "xmm5"
-#endif
+  : "memory", "cc", "xmm0", "xmm1", "xmm4", "xmm5"
   );
 }
 
@@ -631,9 +482,10 @@ void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
   asm volatile (
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    MEMOPREG(pavgb,0x00,0,3,1,xmm0)            //  pavgb   (%0,%3,1),%%xmm0
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,3,1,xmm1)           //  movdqu  (%0,%3,1),%%xmm1
     "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
     "movdqa    %%xmm0,%%xmm1                   \n"
     "pshufb    %%xmm2,%%xmm1                   \n"
     "movdqa    %%xmm0,%%xmm6                   \n"
@@ -643,23 +495,18 @@ void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
     "paddusw   %%xmm0,%%xmm1                   \n"
     "pmulhuw   %%xmm5,%%xmm1                   \n"
     "packuswb  %%xmm1,%%xmm1                   \n"
-    "sub       $0x6,%2                         \n"
     "movd      %%xmm1," MEMACCESS(1) "         \n"
     "psrlq     $0x10,%%xmm1                    \n"
     "movd      %%xmm1," MEMACCESS2(0x2,1) "    \n"
     "lea       " MEMLEA(0x6,1) ",%1            \n"
+    "sub       $0x6,%2                         \n"
     "jg        1b                              \n"
   : "+r"(src_ptr),     // %0
     "+r"(dst_ptr),     // %1
     "+r"(dst_width)    // %2
   : "r"((intptr_t)(src_stride))  // %3
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
   );
 }
 
@@ -679,8 +526,8 @@ void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
   asm volatile (
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    MEMOPREG(movdqa,0x00,0,3,1,xmm6)           //  movdqa  (%0,%3,1),%%xmm6
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,3,1,xmm6)           //  movdqu  (%0,%3,1),%%xmm6
     "movhlps   %%xmm0,%%xmm1                   \n"
     "movhlps   %%xmm6,%%xmm7                   \n"
     "punpcklbw %%xmm5,%%xmm0                   \n"
@@ -689,7 +536,7 @@ void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
     "punpcklbw %%xmm5,%%xmm7                   \n"
     "paddusw   %%xmm6,%%xmm0                   \n"
     "paddusw   %%xmm7,%%xmm1                   \n"
-    MEMOPREG(movdqa,0x00,0,3,2,xmm6)           //  movdqa  (%0,%3,2),%%xmm6
+    MEMOPREG(movdqu,0x00,0,3,2,xmm6)           //  movdqu  (%0,%3,2),%%xmm6
     "lea       " MEMLEA(0x10,0) ",%0           \n"
     "movhlps   %%xmm6,%%xmm7                   \n"
     "punpcklbw %%xmm5,%%xmm6                   \n"
@@ -711,23 +558,18 @@ void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
     "paddusw   %%xmm7,%%xmm6                   \n"
     "pmulhuw   %%xmm4,%%xmm6                   \n"
     "packuswb  %%xmm6,%%xmm6                   \n"
-    "sub       $0x6,%2                         \n"
     "movd      %%xmm6," MEMACCESS(1) "         \n"
     "psrlq     $0x10,%%xmm6                    \n"
     "movd      %%xmm6," MEMACCESS2(0x2,1) "    \n"
     "lea       " MEMLEA(0x6,1) ",%1            \n"
+    "sub       $0x6,%2                         \n"
     "jg        1b                              \n"
   : "+r"(src_ptr),    // %0
     "+r"(dst_ptr),    // %1
     "+r"(dst_width)   // %2
   : "r"((intptr_t)(src_stride))   // %3
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
   );
 }
 
@@ -741,7 +583,7 @@ void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
 
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
     "mov       %0,%3                           \n"
     "add       %6,%0                           \n"
     "movdqa    %%xmm0,%%xmm1                   \n"
@@ -753,7 +595,7 @@ void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
 
     LABELALIGN
   "2:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm2         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
     "add       %6,%0                           \n"
     "movdqa    %%xmm2,%%xmm3                   \n"
     "punpcklbw %%xmm4,%%xmm2                   \n"
@@ -765,8 +607,8 @@ void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
 
     LABELALIGN
   "3:                                          \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
-    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
     "lea       " MEMLEA(0x10,3) ",%0           \n"
     "lea       " MEMLEA(0x20,1) ",%1           \n"
     "sub       $0x10,%4                        \n"
@@ -778,10 +620,7 @@ void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
     "+r"(src_width),   // %4
     "+rm"(src_height)  // %5
   : "rm"((intptr_t)(src_stride))  // %6
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
-#endif
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
   );
 }
 
@@ -813,7 +652,6 @@ void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
     MEMOPARG(movzwl,0x00,1,3,1,k2)             //  movzwl  (%1,%3,1),%k2
     "movd      %k2,%%xmm0                      \n"
     "psrlw     $0x9,%%xmm1                     \n"
-    BUNDLEALIGN
     MEMOPARG(movzwl,0x00,1,4,1,k2)             //  movzwl  (%1,%4,1),%k2
     "movd      %k2,%%xmm4                      \n"
     "pshufb    %%xmm5,%%xmm1                   \n"
@@ -853,13 +691,8 @@ void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
     "+rm"(dst_width)   // %5
   : "rm"(x),           // %6
     "rm"(dx)           // %7
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
   );
 }
 
@@ -870,25 +703,21 @@ void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr,
   asm volatile (
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
     "lea       " MEMLEA(0x10,1) ",%1           \n"
     "movdqa    %%xmm0,%%xmm1                   \n"
     "punpcklbw %%xmm0,%%xmm0                   \n"
     "punpckhbw %%xmm1,%%xmm1                   \n"
-    "sub       $0x20,%2                         \n"
-    "movdqa    %%xmm0," MEMACCESS(0) "         \n"
-    "movdqa    %%xmm1," MEMACCESS2(0x10,0) "   \n"
+    "movdqu    %%xmm0," MEMACCESS(0) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,0) "   \n"
     "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "sub       $0x20,%2                         \n"
     "jg        1b                              \n"
 
   : "+r"(dst_ptr),     // %0
     "+r"(src_ptr),     // %1
     "+r"(dst_width)    // %2
-  :
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1"
-#endif
+  :: "memory", "cc", "xmm0", "xmm1"
   );
 }
 
@@ -898,22 +727,18 @@ void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
   asm volatile (
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
     "lea       " MEMLEA(0x20,0) ",%0           \n"
     "shufps    $0xdd,%%xmm1,%%xmm0             \n"
-    "sub       $0x4,%2                         \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%2                         \n"
     "jg        1b                              \n"
   : "+r"(src_argb),  // %0
     "+r"(dst_argb),  // %1
     "+r"(dst_width)  // %2
-  :
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1"
-#endif
+  :: "memory", "cc", "xmm0", "xmm1"
   );
 }
 
@@ -923,25 +748,21 @@ void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb,
   asm volatile (
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
     "lea       " MEMLEA(0x20,0) ",%0           \n"
     "movdqa    %%xmm0,%%xmm2                   \n"
     "shufps    $0x88,%%xmm1,%%xmm0             \n"
     "shufps    $0xdd,%%xmm1,%%xmm2             \n"
     "pavgb     %%xmm2,%%xmm0                   \n"
-    "sub       $0x4,%2                         \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%2                         \n"
     "jg        1b                              \n"
   : "+r"(src_argb),  // %0
     "+r"(dst_argb),  // %1
     "+r"(dst_width)  // %2
-  :
-  : "memory", "cc"
-#if defined(__SSE2__)
-    , "xmm0", "xmm1"
-#endif
+  :: "memory", "cc", "xmm0", "xmm1"
   );
 }
 
@@ -951,11 +772,10 @@ void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
   asm volatile (
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    BUNDLEALIGN
-    MEMOPREG(movdqa,0x00,0,3,1,xmm2)           //  movdqa   (%0,%3,1),%%xmm2
-    MEMOPREG(movdqa,0x10,0,3,1,xmm3)           //  movdqa   0x10(%0,%3,1),%%xmm3
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x00,0,3,1,xmm2)           //  movdqu   (%0,%3,1),%%xmm2
+    MEMOPREG(movdqu,0x10,0,3,1,xmm3)           //  movdqu   0x10(%0,%3,1),%%xmm3
     "lea       " MEMLEA(0x20,0) ",%0           \n"
     "pavgb     %%xmm2,%%xmm0                   \n"
     "pavgb     %%xmm3,%%xmm1                   \n"
@@ -963,21 +783,16 @@ void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
     "shufps    $0x88,%%xmm1,%%xmm0             \n"
     "shufps    $0xdd,%%xmm1,%%xmm2             \n"
     "pavgb     %%xmm2,%%xmm0                   \n"
-    "sub       $0x4,%2                         \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
     "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%2                         \n"
     "jg        1b                              \n"
   : "+r"(src_argb),   // %0
     "+r"(dst_argb),   // %1
     "+r"(dst_width)   // %2
   : "r"((intptr_t)(src_stride))   // %3
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3"
   );
 }
 
@@ -996,29 +811,22 @@ void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
     "movd      " MEMACCESS(0) ",%%xmm0         \n"
     MEMOPREG(movd,0x00,0,1,1,xmm1)             //  movd      (%0,%1,1),%%xmm1
     "punpckldq %%xmm1,%%xmm0                   \n"
-    BUNDLEALIGN
     MEMOPREG(movd,0x00,0,1,2,xmm2)             //  movd      (%0,%1,2),%%xmm2
     MEMOPREG(movd,0x00,0,4,1,xmm3)             //  movd      (%0,%4,1),%%xmm3
     "lea       " MEMLEA4(0x00,0,1,4) ",%0      \n"
     "punpckldq %%xmm3,%%xmm2                   \n"
     "punpcklqdq %%xmm2,%%xmm0                  \n"
-    "sub       $0x4,%3                         \n"
-    "movdqa    %%xmm0," MEMACCESS(2) "         \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
     "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
     "jg        1b                              \n"
   : "+r"(src_argb),      // %0
     "+r"(src_stepx_x4),  // %1
     "+r"(dst_argb),      // %2
     "+r"(dst_width),     // %3
     "+r"(src_stepx_x12)  // %4
-  :
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3"
-#endif
+  :: "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3"
   );
 }
 
@@ -1040,11 +848,9 @@ void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
     "movq      " MEMACCESS(0) ",%%xmm0         \n"
     MEMOPREG(movhps,0x00,0,1,1,xmm0)           //  movhps    (%0,%1,1),%%xmm0
     MEMOPREG(movq,0x00,0,1,2,xmm1)             //  movq      (%0,%1,2),%%xmm1
-    BUNDLEALIGN
     MEMOPREG(movhps,0x00,0,4,1,xmm1)           //  movhps    (%0,%4,1),%%xmm1
     "lea       " MEMLEA4(0x00,0,1,4) ",%0      \n"
     "movq      " MEMACCESS(5) ",%%xmm2         \n"
-    BUNDLEALIGN
     MEMOPREG(movhps,0x00,5,1,1,xmm2)           //  movhps    (%5,%1,1),%%xmm2
     MEMOPREG(movq,0x00,5,1,2,xmm3)             //  movq      (%5,%1,2),%%xmm3
     MEMOPREG(movhps,0x00,5,4,1,xmm3)           //  movhps    (%5,%4,1),%%xmm3
@@ -1055,9 +861,9 @@ void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
     "shufps    $0x88,%%xmm1,%%xmm0             \n"
     "shufps    $0xdd,%%xmm1,%%xmm2             \n"
     "pavgb     %%xmm2,%%xmm0                   \n"
-    "sub       $0x4,%3                         \n"
-    "movdqa    %%xmm0," MEMACCESS(2) "         \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
     "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
     "jg        1b                              \n"
   : "+r"(src_argb),       // %0
     "+r"(src_stepx_x4),   // %1
@@ -1065,14 +871,8 @@ void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
     "+rm"(dst_width),     // %3
     "+r"(src_stepx_x12),  // %4
     "+r"(row1)            // %5
-  :
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3"
-#endif
+  :: "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3"
   );
 }
 
@@ -1111,15 +911,14 @@ void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
     "pextrw    $0x3,%%xmm2,%k1                 \n"
     "punpckldq %%xmm4,%%xmm1                   \n"
     "punpcklqdq %%xmm1,%%xmm0                  \n"
-    "sub       $0x4,%4                         \n"
     "movdqu    %%xmm0," MEMACCESS(2) "         \n"
     "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%4                         \n"
     "jge       40b                             \n"
 
   "49:                                         \n"
     "test      $0x2,%4                         \n"
     "je        29f                             \n"
-    BUNDLEALIGN
     MEMOPREG(movd,0x00,3,0,4,xmm0)             //  movd      (%3,%0,4),%%xmm0
     MEMOPREG(movd,0x00,3,1,4,xmm1)             //  movd      (%3,%1,4),%%xmm1
     "pextrw    $0x5,%%xmm2,%k0                 \n"
@@ -1139,13 +938,8 @@ void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
     "+r"(dst_width)    // %4
   : "rm"(x),           // %5
     "rm"(dx)           // %6
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
   );
 }
 
@@ -1156,28 +950,22 @@ void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
   asm volatile (
     LABELALIGN
   "1:                                          \n"
-    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
     "lea       " MEMLEA(0x10,1) ",%1           \n"
     "movdqa    %%xmm0,%%xmm1                   \n"
     "punpckldq %%xmm0,%%xmm0                   \n"
     "punpckhdq %%xmm1,%%xmm1                   \n"
-    "sub       $0x8,%2                         \n"
-    "movdqa    %%xmm0," MEMACCESS(0) "         \n"
-    "movdqa    %%xmm1," MEMACCESS2(0x10,0) "   \n"
+    "movdqu    %%xmm0," MEMACCESS(0) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,0) "   \n"
     "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "sub       $0x8,%2                         \n"
     "jg        1b                              \n"
 
   : "+r"(dst_argb),    // %0
     "+r"(src_argb),    // %1
     "+r"(dst_width)    // %2
-  :
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1"
-#endif
+  :: "memory", "cc", NACL_R14
+    "xmm0", "xmm1"
   );
 }
 
@@ -1225,7 +1013,6 @@ void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
     "paddd     %%xmm3,%%xmm2                   \n"
     MEMOPREG(movq,0x00,1,3,4,xmm0)             //  movq      (%1,%3,4),%%xmm0
     "psrlw     $0x9,%%xmm1                     \n"
-    BUNDLEALIGN
     MEMOPREG(movhps,0x00,1,4,4,xmm0)           //  movhps    (%1,%4,4),%%xmm0
     "pshufb    %%xmm5,%%xmm1                   \n"
     "pshufb    %%xmm4,%%xmm0                   \n"
@@ -1245,7 +1032,6 @@ void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
     "add       $0x1,%2                         \n"
     "jl        99f                             \n"
     "psrlw     $0x9,%%xmm2                     \n"
-    BUNDLEALIGN
     MEMOPREG(movq,0x00,1,3,4,xmm0)             //  movq      (%1,%3,4),%%xmm0
     "pshufb    %%xmm5,%%xmm2                   \n"
     "pshufb    %%xmm4,%%xmm0                   \n"
@@ -1264,13 +1050,8 @@ void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
     "+r"(x1)           // %4
   : "rm"(x),           // %5
     "rm"(dx)           // %6
-  : "memory", "cc"
-#if defined(__native_client__) && defined(__x86_64__)
-    , "r14"
-#endif
-#if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
-#endif
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
   );
 }
 
diff --git a/third_party/libyuv/source/scale_win.cc b/third_party/libyuv/source/scale_win.cc
index 840b9738d..e0209cdec 100644
--- a/third_party/libyuv/source/scale_win.cc
+++ b/third_party/libyuv/source/scale_win.cc
@@ -103,118 +103,6 @@ void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
     mov        edx, [esp + 12]       // dst_ptr
     mov        ecx, [esp + 16]       // dst_width
 
-    align      4
-  wloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    lea        eax,  [eax + 32]
-    psrlw      xmm0, 8               // isolate odd pixels.
-    psrlw      xmm1, 8
-    packuswb   xmm0, xmm1
-    sub        ecx, 16
-    movdqa     [edx], xmm0
-    lea        edx, [edx + 16]
-    jg         wloop
-
-    ret
-  }
-}
-
-// Blends 32x1 rectangle to 16x1.
-// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
-__declspec(naked) __declspec(align(16))
-void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
-                              uint8* dst_ptr, int dst_width) {
-  __asm {
-    mov        eax, [esp + 4]        // src_ptr
-                                     // src_stride
-    mov        edx, [esp + 12]       // dst_ptr
-    mov        ecx, [esp + 16]       // dst_width
-    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
-    psrlw      xmm5, 8
-
-    align      4
-  wloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    lea        eax,  [eax + 32]
-
-    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
-    psrlw      xmm0, 8
-    movdqa     xmm3, xmm1
-    psrlw      xmm1, 8
-    pand       xmm2, xmm5
-    pand       xmm3, xmm5
-    pavgw      xmm0, xmm2
-    pavgw      xmm1, xmm3
-    packuswb   xmm0, xmm1
-
-    sub        ecx, 16
-    movdqa     [edx], xmm0
-    lea        edx, [edx + 16]
-    jg         wloop
-
-    ret
-  }
-}
-
-// Blends 32x2 rectangle to 16x1.
-// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
-__declspec(naked) __declspec(align(16))
-void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
-                           uint8* dst_ptr, int dst_width) {
-  __asm {
-    push       esi
-    mov        eax, [esp + 4 + 4]    // src_ptr
-    mov        esi, [esp + 4 + 8]    // src_stride
-    mov        edx, [esp + 4 + 12]   // dst_ptr
-    mov        ecx, [esp + 4 + 16]   // dst_width
-    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
-    psrlw      xmm5, 8
-
-    align      4
-  wloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    movdqa     xmm2, [eax + esi]
-    movdqa     xmm3, [eax + esi + 16]
-    lea        eax,  [eax + 32]
-    pavgb      xmm0, xmm2            // average rows
-    pavgb      xmm1, xmm3
-
-    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
-    psrlw      xmm0, 8
-    movdqa     xmm3, xmm1
-    psrlw      xmm1, 8
-    pand       xmm2, xmm5
-    pand       xmm3, xmm5
-    pavgw      xmm0, xmm2
-    pavgw      xmm1, xmm3
-    packuswb   xmm0, xmm1
-
-    sub        ecx, 16
-    movdqa     [edx], xmm0
-    lea        edx, [edx + 16]
-    jg         wloop
-
-    pop        esi
-    ret
-  }
-}
-
-// Reads 32 pixels, throws half away and writes 16 pixels.
-// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
-__declspec(naked) __declspec(align(16))
-void ScaleRowDown2_Unaligned_SSE2(const uint8* src_ptr,
-                                  ptrdiff_t src_stride,
-                                  uint8* dst_ptr, int dst_width) {
-  __asm {
-    mov        eax, [esp + 4]        // src_ptr
-                                     // src_stride ignored
-    mov        edx, [esp + 12]       // dst_ptr
-    mov        ecx, [esp + 16]       // dst_width
-
-    align      4
   wloop:
     movdqu     xmm0, [eax]
     movdqu     xmm1, [eax + 16]
@@ -222,9 +110,9 @@ void ScaleRowDown2_Unaligned_SSE2(const uint8* src_ptr,
     psrlw      xmm0, 8               // isolate odd pixels.
     psrlw      xmm1, 8
     packuswb   xmm0, xmm1
-    sub        ecx, 16
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 16
     jg         wloop
 
     ret
@@ -234,9 +122,8 @@ void ScaleRowDown2_Unaligned_SSE2(const uint8* src_ptr,
 // Blends 32x1 rectangle to 16x1.
 // Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
 __declspec(naked) __declspec(align(16))
-void ScaleRowDown2Linear_Unaligned_SSE2(const uint8* src_ptr,
-                                        ptrdiff_t src_stride,
-                                        uint8* dst_ptr, int dst_width) {
+void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width) {
   __asm {
     mov        eax, [esp + 4]        // src_ptr
                                      // src_stride
@@ -245,7 +132,6 @@ void ScaleRowDown2Linear_Unaligned_SSE2(const uint8* src_ptr,
     pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
     psrlw      xmm5, 8
 
-    align      4
   wloop:
     movdqu     xmm0, [eax]
     movdqu     xmm1, [eax + 16]
@@ -261,9 +147,9 @@ void ScaleRowDown2Linear_Unaligned_SSE2(const uint8* src_ptr,
     pavgw      xmm1, xmm3
     packuswb   xmm0, xmm1
 
-    sub        ecx, 16
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 16
     jg         wloop
 
     ret
@@ -273,9 +159,8 @@ void ScaleRowDown2Linear_Unaligned_SSE2(const uint8* src_ptr,
 // Blends 32x2 rectangle to 16x1.
 // Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
 __declspec(naked) __declspec(align(16))
-void ScaleRowDown2Box_Unaligned_SSE2(const uint8* src_ptr,
-                                     ptrdiff_t src_stride,
-                                     uint8* dst_ptr, int dst_width) {
+void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
   __asm {
     push       esi
     mov        eax, [esp + 4 + 4]    // src_ptr
@@ -285,7 +170,6 @@ void ScaleRowDown2Box_Unaligned_SSE2(const uint8* src_ptr,
     pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
     psrlw      xmm5, 8
 
-    align      4
   wloop:
     movdqu     xmm0, [eax]
     movdqu     xmm1, [eax + 16]
@@ -305,9 +189,9 @@ void ScaleRowDown2Box_Unaligned_SSE2(const uint8* src_ptr,
     pavgw      xmm1, xmm3
     packuswb   xmm0, xmm1
 
-    sub        ecx, 16
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 16
     jg         wloop
 
     pop        esi
@@ -329,19 +213,18 @@ void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
     psrld      xmm5, 24
     pslld      xmm5, 16
 
-    align      4
   wloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
     lea        eax,  [eax + 32]
     pand       xmm0, xmm5
     pand       xmm1, xmm5
     packuswb   xmm0, xmm1
     psrlw      xmm0, 8
     packuswb   xmm0, xmm0
-    sub        ecx, 8
     movq       qword ptr [edx], xmm0
     lea        edx, [edx + 8]
+    sub        ecx, 8
     jg         wloop
 
     ret
@@ -364,18 +247,17 @@ void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
     pcmpeqb    xmm7, xmm7            // generate mask 0x00ff00ff
     psrlw      xmm7, 8
 
-    align      4
   wloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    movdqa     xmm2, [eax + esi]
-    movdqa     xmm3, [eax + esi + 16]
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + esi]
+    movdqu     xmm3, [eax + esi + 16]
     pavgb      xmm0, xmm2            // average rows
     pavgb      xmm1, xmm3
-    movdqa     xmm2, [eax + esi * 2]
-    movdqa     xmm3, [eax + esi * 2 + 16]
-    movdqa     xmm4, [eax + edi]
-    movdqa     xmm5, [eax + edi + 16]
+    movdqu     xmm2, [eax + esi * 2]
+    movdqu     xmm3, [eax + esi * 2 + 16]
+    movdqu     xmm4, [eax + edi]
+    movdqu     xmm5, [eax + edi + 16]
     lea        eax, [eax + 32]
     pavgb      xmm2, xmm4
     pavgb      xmm3, xmm5
@@ -398,9 +280,9 @@ void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
     pavgw      xmm0, xmm2
     packuswb   xmm0, xmm0
 
-    sub        ecx, 8
     movq       qword ptr [edx], xmm0
     lea        edx, [edx + 8]
+    sub        ecx, 8
     jg         wloop
 
     pop        edi
@@ -427,10 +309,9 @@ void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
     movdqa     xmm4, kShuf1
     movdqa     xmm5, kShuf2
 
-    align      4
   wloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
     lea        eax,  [eax + 32]
     movdqa     xmm2, xmm1
     palignr    xmm1, xmm0, 8
@@ -481,10 +362,9 @@ void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
     movdqa     xmm6, kMadd11
     movdqa     xmm7, kRound34
 
-    align      4
   wloop:
-    movdqa     xmm0, [eax]           // pixels 0..7
-    movdqa     xmm1, [eax + esi]
+    movdqu     xmm0, [eax]           // pixels 0..7
+    movdqu     xmm1, [eax + esi]
     pavgb      xmm0, xmm1
     pshufb     xmm0, xmm2
     pmaddubsw  xmm0, xmm5
@@ -501,8 +381,8 @@ void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
     psrlw      xmm0, 2
     packuswb   xmm0, xmm0
     movq       qword ptr [edx + 8], xmm0
-    movdqa     xmm0, [eax + 16]      // pixels 16..23
-    movdqa     xmm1, [eax + esi + 16]
+    movdqu     xmm0, [eax + 16]      // pixels 16..23
+    movdqu     xmm1, [eax + esi + 16]
     lea        eax, [eax + 32]
     pavgb      xmm0, xmm1
     pshufb     xmm0, xmm4
@@ -511,9 +391,9 @@ void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
     paddsw     xmm0, xmm7
     psrlw      xmm0, 2
     packuswb   xmm0, xmm0
-    sub        ecx, 24
     movq       qword ptr [edx + 16], xmm0
     lea        edx, [edx + 24]
+    sub        ecx, 24
     jg         wloop
 
     pop        esi
@@ -540,10 +420,9 @@ void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
     movdqa     xmm6, kMadd11
     movdqa     xmm7, kRound34
 
-    align      4
   wloop:
-    movdqa     xmm0, [eax]           // pixels 0..7
-    movdqa     xmm1, [eax + esi]
+    movdqu     xmm0, [eax]           // pixels 0..7
+    movdqu     xmm1, [eax + esi]
     pavgb      xmm1, xmm0
     pavgb      xmm0, xmm1
     pshufb     xmm0, xmm2
@@ -562,8 +441,8 @@ void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
     psrlw      xmm0, 2
     packuswb   xmm0, xmm0
     movq       qword ptr [edx + 8], xmm0
-    movdqa     xmm0, [eax + 16]      // pixels 16..23
-    movdqa     xmm1, [eax + esi + 16]
+    movdqu     xmm0, [eax + 16]      // pixels 16..23
+    movdqu     xmm1, [eax + esi + 16]
     lea        eax, [eax + 32]
     pavgb      xmm1, xmm0
     pavgb      xmm0, xmm1
@@ -573,9 +452,9 @@ void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
     paddsw     xmm0, xmm7
     psrlw      xmm0, 2
     packuswb   xmm0, xmm0
-    sub        ecx, 24
     movq       qword ptr [edx + 16], xmm0
     lea        edx, [edx+24]
+    sub        ecx, 24
     jg         wloop
 
     pop        esi
@@ -597,20 +476,19 @@ void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
     movdqa     xmm4, kShuf38a
     movdqa     xmm5, kShuf38b
 
-    align      4
   xloop:
-    movdqa     xmm0, [eax]           // 16 pixels -> 0,1,2,3,4,5
-    movdqa     xmm1, [eax + 16]      // 16 pixels -> 6,7,8,9,10,11
+    movdqu     xmm0, [eax]           // 16 pixels -> 0,1,2,3,4,5
+    movdqu     xmm1, [eax + 16]      // 16 pixels -> 6,7,8,9,10,11
     lea        eax, [eax + 32]
     pshufb     xmm0, xmm4
     pshufb     xmm1, xmm5
     paddusb    xmm0, xmm1
 
-    sub        ecx, 12
     movq       qword ptr [edx], xmm0  // write 12 pixels
     movhlps    xmm1, xmm0
     movd       [edx + 8], xmm1
     lea        edx, [edx + 12]
+    sub        ecx, 12
     jg         xloop
 
     ret
@@ -633,10 +511,9 @@ void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
     movdqa     xmm4, kScaleAc33
     pxor       xmm5, xmm5
 
-    align      4
   xloop:
-    movdqa     xmm0, [eax]           // sum up 3 rows into xmm0/1
-    movdqa     xmm6, [eax + esi]
+    movdqu     xmm0, [eax]           // sum up 3 rows into xmm0/1
+    movdqu     xmm6, [eax + esi]
     movhlps    xmm1, xmm0
     movhlps    xmm7, xmm6
     punpcklbw  xmm0, xmm5
@@ -645,7 +522,7 @@ void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
     punpcklbw  xmm7, xmm5
     paddusw    xmm0, xmm6
     paddusw    xmm1, xmm7
-    movdqa     xmm6, [eax + esi * 2]
+    movdqu     xmm6, [eax + esi * 2]
     lea        eax, [eax + 16]
     movhlps    xmm7, xmm6
     punpcklbw  xmm6, xmm5
@@ -671,11 +548,11 @@ void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
     pmulhuw    xmm6, xmm4            // divide by 9,9,6, 9,9,6
     packuswb   xmm6, xmm6
 
-    sub        ecx, 6
     movd       [edx], xmm6           // write 6 pixels
     psrlq      xmm6, 16
     movd       [edx + 2], xmm6
     lea        edx, [edx + 6]
+    sub        ecx, 6
     jg         xloop
 
     pop        esi
@@ -699,11 +576,11 @@ void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
     movdqa     xmm4, kShufAb2
     movdqa     xmm5, kScaleAb2
 
-    align      4
   xloop:
-    movdqa     xmm0, [eax]           // average 2 rows into xmm0
-    pavgb      xmm0, [eax + esi]
+    movdqu     xmm0, [eax]           // average 2 rows into xmm0
+    movdqu     xmm1, [eax + esi]
     lea        eax, [eax + 16]
+    pavgb      xmm0, xmm1
 
     movdqa     xmm1, xmm0            // 16 pixels -> 0,1,2,3,4,5 of xmm1
     pshufb     xmm1, xmm2
@@ -716,11 +593,11 @@ void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
     pmulhuw    xmm1, xmm5            // divide by 3,3,2, 3,3,2
     packuswb   xmm1, xmm1
 
-    sub        ecx, 6
     movd       [edx], xmm1           // write 6 pixels
     psrlq      xmm1, 16
     movd       [edx + 2], xmm1
     lea        edx, [edx + 6]
+    sub        ecx, 6
     jg         xloop
 
     pop        esi
@@ -747,10 +624,9 @@ void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
     pxor       xmm4, xmm4
     dec        ebx
 
-    align      4
   xloop:
     // first row
-    movdqa     xmm0, [esi]
+    movdqu     xmm0, [esi]
     lea        eax, [esi + edx]
     movdqa     xmm1, xmm0
     punpcklbw  xmm0, xmm4
@@ -761,9 +637,8 @@ void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
     je         ydone
 
     // sum remaining rows
-    align      4
   yloop:
-    movdqa     xmm2, [eax]       // read 16 pixels
+    movdqu     xmm2, [eax]       // read 16 pixels
     lea        eax, [eax + edx]  // advance to next row
     movdqa     xmm3, xmm2
     punpcklbw  xmm2, xmm4
@@ -773,10 +648,9 @@ void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
     sub        ebp, 1
     jg         yloop
 
-    align      4
   ydone:
-    movdqa     [edi], xmm0
-    movdqa     [edi + 16], xmm1
+    movdqu     [edi], xmm0
+    movdqu     [edi + 16], xmm1
     lea        edi, [edi + 32]
 
     sub        ecx, 16
@@ -828,7 +702,6 @@ void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
     pextrw     edx, xmm2, 3         // get x1 integer. preroll
 
     // 2 Pixel loop.
-    align      4
   xloop2:
     movdqa     xmm1, xmm2           // x0, x1 fractions.
     paddd      xmm2, xmm3           // x += dx
@@ -851,7 +724,6 @@ void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
     sub        ecx, 2               // 2 pixels
     jge        xloop2
 
-    align      4
  xloop29:
 
     add        ecx, 2 - 1
@@ -869,7 +741,6 @@ void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
     movd       ebx, xmm0
     mov        [edi], bl
 
-    align      4
  xloop99:
 
     pop        edi
@@ -889,17 +760,16 @@ void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr,
     mov        eax, [esp + 8]    // src_ptr
     mov        ecx, [esp + 12]   // dst_width
 
-    align      4
   wloop:
-    movdqa     xmm0, [eax]
+    movdqu     xmm0, [eax]
     lea        eax,  [eax + 16]
     movdqa     xmm1, xmm0
     punpcklbw  xmm0, xmm0
     punpckhbw  xmm1, xmm1
-    sub        ecx, 32
-    movdqa     [edx], xmm0
-    movdqa     [edx + 16], xmm1
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
     lea        edx, [edx + 32]
+    sub        ecx, 32
     jg         wloop
 
     ret
@@ -918,15 +788,14 @@ void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
     mov        edx, [esp + 12]       // dst_argb
     mov        ecx, [esp + 16]       // dst_width
 
-    align      4
   wloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
     lea        eax,  [eax + 32]
     shufps     xmm0, xmm1, 0xdd
-    sub        ecx, 4
-    movdqa     [edx], xmm0
+    movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 4
     jg         wloop
 
     ret
@@ -945,18 +814,17 @@ void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb,
     mov        edx, [esp + 12]       // dst_argb
     mov        ecx, [esp + 16]       // dst_width
 
-    align      4
   wloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
     lea        eax,  [eax + 32]
     movdqa     xmm2, xmm0
     shufps     xmm0, xmm1, 0x88      // even pixels
     shufps     xmm2, xmm1, 0xdd      // odd pixels
     pavgb      xmm0, xmm2
-    sub        ecx, 4
-    movdqa     [edx], xmm0
+    movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 4
     jg         wloop
 
     ret
@@ -976,12 +844,11 @@ void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
     mov        edx, [esp + 4 + 12]   // dst_argb
     mov        ecx, [esp + 4 + 16]   // dst_width
 
-    align      4
   wloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    movdqa     xmm2, [eax + esi]
-    movdqa     xmm3, [eax + esi + 16]
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + esi]
+    movdqu     xmm3, [eax + esi + 16]
     lea        eax,  [eax + 32]
     pavgb      xmm0, xmm2            // average rows
     pavgb      xmm1, xmm3
@@ -989,9 +856,9 @@ void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
     shufps     xmm0, xmm1, 0x88      // even pixels
     shufps     xmm2, xmm1, 0xdd      // odd pixels
     pavgb      xmm0, xmm2
-    sub        ecx, 4
-    movdqa     [edx], xmm0
+    movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 4
     jg         wloop
 
     pop        esi
@@ -1016,7 +883,6 @@ void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
     lea        ebx, [ebx * 4]
     lea        edi, [ebx + ebx * 2]
 
-    align      4
   wloop:
     movd       xmm0, [eax]
     movd       xmm1, [eax + ebx]
@@ -1026,9 +892,9 @@ void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
     lea        eax,  [eax + ebx * 4]
     punpckldq  xmm2, xmm3
     punpcklqdq xmm0, xmm2
-    sub        ecx, 4
-    movdqa     [edx], xmm0
+    movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 4
     jg         wloop
 
     pop        edi
@@ -1057,7 +923,6 @@ void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
     lea        ebx, [ebx * 4]
     lea        edi, [ebx + ebx * 2]
 
-    align      4
   wloop:
     movq       xmm0, qword ptr [eax]  // row0 4 pairs
     movhps     xmm0, qword ptr [eax + ebx]
@@ -1075,9 +940,9 @@ void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
     shufps     xmm0, xmm1, 0x88      // even pixels
     shufps     xmm2, xmm1, 0xdd      // odd pixels
     pavgb      xmm0, xmm2
-    sub        ecx, 4
-    movdqa     [edx], xmm0
+    movdqu     [edx], xmm0
     lea        edx, [edx + 16]
+    sub        ecx, 4
     jg         wloop
 
     pop        edi
@@ -1118,7 +983,6 @@ void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
     jl         xloop49
 
     // 4 Pixel loop.
-    align      4
  xloop4:
     movd       xmm0, [esi + eax * 4]  // 1 source x0 pixels
     movd       xmm1, [esi + edx * 4]  // 1 source x1 pixels
@@ -1133,12 +997,11 @@ void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
     pextrw     edx, xmm2, 3           // get x1 integer. next iteration.
     punpckldq  xmm1, xmm4             // x2 x3
     punpcklqdq xmm0, xmm1             // x0 x1 x2 x3
-    sub        ecx, 4                 // 4 pixels
     movdqu     [edi], xmm0
     lea        edi, [edi + 16]
+    sub        ecx, 4                 // 4 pixels
     jge        xloop4
 
-    align      4
  xloop49:
     test       ecx, 2
     je         xloop29
@@ -1159,7 +1022,6 @@ void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
     // 1 Pixels.
     movd       xmm0, [esi + eax * 4]  // 1 source x2 pixels
     movd       dword ptr [edi], xmm0
-    align      4
  xloop99:
 
     pop        esi
@@ -1209,7 +1071,6 @@ void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
     pextrw     edx, xmm2, 3         // get x1 integer. preroll
 
     // 2 Pixel loop.
-    align      4
   xloop2:
     movdqa     xmm1, xmm2           // x0, x1 fractions.
     paddd      xmm2, xmm3           // x += dx
@@ -1229,7 +1090,6 @@ void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
     sub        ecx, 2               // 2 pixels
     jge        xloop2
 
-    align      4
  xloop29:
 
     add        ecx, 2 - 1
@@ -1246,7 +1106,6 @@ void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
     packuswb   xmm0, xmm0           // argb 8 bits, 1 pixel.
     movd       [edi], xmm0
 
-    align      4
  xloop99:
 
     pop        edi
@@ -1265,17 +1124,16 @@ void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
     mov        eax, [esp + 8]    // src_argb
     mov        ecx, [esp + 12]   // dst_width
 
-    align      4
   wloop:
-    movdqa     xmm0, [eax]
+    movdqu     xmm0, [eax]
     lea        eax,  [eax + 16]
     movdqa     xmm1, xmm0
     punpckldq  xmm0, xmm0
     punpckhdq  xmm1, xmm1
-    sub        ecx, 8
-    movdqa     [edx], xmm0
-    movdqa     [edx + 16], xmm1
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
     lea        edx, [edx + 32]
+    sub        ecx, 8
     jg         wloop
 
     ret
diff --git a/third_party/libyuv/source/video_common.cc b/third_party/libyuv/source/video_common.cc
index efbedf46e..379a0669a 100644
--- a/third_party/libyuv/source/video_common.cc
+++ b/third_party/libyuv/source/video_common.cc
@@ -33,7 +33,7 @@ static const struct FourCCAliasEntry kFourCCAliases[] = {
   {FOURCC_2VUY, FOURCC_UYVY},  // kCMPixelFormat_422YpCbCr8
   {FOURCC_JPEG, FOURCC_MJPG},  // Note: JPEG has DHT while MJPG does not.
   {FOURCC_DMB1, FOURCC_MJPG},
-  {FOURCC_BA81, FOURCC_BGGR},
+  {FOURCC_BA81, FOURCC_BGGR},  // deprecated.
   {FOURCC_RGB3, FOURCC_RAW },
   {FOURCC_BGR3, FOURCC_24BG},
   {FOURCC_CM32, FOURCC_BGRA},  // kCMPixelFormat_32ARGB
author	Johann <johannkoenig@google.com>	2015-03-02 15:19:19 -0800
committer	Johann <johannkoenig@google.com>	2015-03-02 15:19:31 -0800
commit	223bf29307a1981940dbc224a0b1279a187eca6c (patch)
tree	438d592b79c8f70ef5ea5c0c83cc1c02484c19d3 /third_party/libyuv/source
parent	1790d45252efb29903baae3d1776bb24cee76558 (diff)
download	libvpx-223bf29307a1981940dbc224a0b1279a187eca6c.tar libvpx-223bf29307a1981940dbc224a0b1279a187eca6c.tar.gz libvpx-223bf29307a1981940dbc224a0b1279a187eca6c.tar.bz2 libvpx-223bf29307a1981940dbc224a0b1279a187eca6c.zip